Taxpayer Compliance, Volume 1: An Agenda for Research [Reprint 2016 ed.] 9781512806274

Table of contents :
Contents
Preface
Summary
1. Paying Taxes
2. Understanding Taxpayer Compliance: Self-Interest, Social Commitment, and Other Influences
3. Expanding the Framework of Analysis
4. Extending Research on Tax Administration
5. Data Needs for Taxpayer Compliance Research
6. Getting Started: What Needs to be Done
References and Bibliography
Appendices
Appendix A: Statistical Issues in Modeling Taxpayer Compliance
Appendix B: Experimental and Quasi-Experimental Designs in Taxpayer Compliance Research
Appendix C: Symposium on Taxpayer Compliance Research
Appendix D: Panel on Taxpayer Compliance Research
Appendix E: Committee on Law Enforcement and the Administration of Justice, 1987-1988
Index

Citation preview

Taxpayer Compliance Volume i: An Agenda for Research

Law in Social Context Series EDITORS:

Keith Hawkins, Oxford University, Centre for Socio-Legal Studies JohnM. Thomas, State University of New York at Buffalo, School of Management

Taxpayer Compliance Volume i: An Agenda for Research Jeffrey A. Roth, John T. Scholz, and Ann Dryden Witte, Editors Panel on Taxpayer Compliance Research Committee on Research on Law Enforcement and the Administration of Justice Commission on Behavioral and Social Sciences and Education National Research Council

Ufll University of Pennsylvania Press Philadelphia

NOTICE: The project that is the subject of this report was approved by the Governing Board of the National Research Council, whose members are drawn from the councils of the National Academy of Sciences, the National Academy of Engineering, and the Institute of Medicine. The members of the committee responsible for the report were chosen for their special competences and with regard for appropriate balance. This report has been reviewed by a group other than the authors according to procedures approved by a Report Review Committee consisting of members of the National Academy of Sciences, the National Academy of Engineering, and the Institute of Medicine. The National Research Council was established by the National Academy of Sciences in 1916 to associate the broad community of science and technology with the Academy's purposes of furthering knowledge and of advising the federal government. The Council operates in accordance with general policies determined by the Academy under the authority of its congressional charter of 1863, which establishes the Academy as a private, nonprofit, self-governing membership corporation. The Council has become the principal operating agency of both the National Academy of Sciences and the National Academy of Engineering in the conduct of their services to the government, the public, and the scientific and engineering communities. It is administered jointly by both Academies and the Institute of Medicine. The National Academy of Engineering and the Institute of Medicine were established in 1964 and 1970, respectively, under the charter of the National Academy of Sciences. This project was sponsored by the Internal Revenue Service, U.S. Department of the Treasury, under contract no. 84-0092. The contents do not necessarily reflect the views or policies of the sponsor.

Copyright © 1989 by the University of Pennnsylvania Press All rights reserved Printed in the United States of America Library of Congress Cataloging in Publication Data Taypayer compliance. (Law in social context series) Bibliography: p. Includes index. Contents: v. 1. A n agenda for research— v. 2. Social science perspectives. 1. Taxpayer compliance—United States. 2. Taxation— L a w and legislation—United States. I. Roth, Jeffrey A., 1945II. Scholz, John T. III. Witte, Ann D . I V . Series. HJ4652.T396 1989 3?6.2'9i 88-56250 I S B N 0-8122-8182-9 (v. 1) I S B N 0-8122-8150-0 (v. 2) I S B N 0-8122-8187-X (set)

Contents

Preface Summary

1

2

Paying Taxes Supplementary Notes A: Patterns and Trends in Taxpayer Compliance B: The Taxpayer Compliance Measurement Program

vii i

16 40 65

Understanding Taxpayer Compliance: Self-interest, Social Commitment, and Other Influences

71

3 Expanding the Framework of Analysis

142

4

Extending Research on Tax Administration

180

5 Data Needs for Taxpayer Compliance Research

207

6

247

Getting Started: What Needs to be Done

References and Bibliography

265

Appendices A

Statistical Issues in Modeling Taxpayer Compliance Peter Schmidt

307

B

Experimental and Quasi-Experimental Designs in Taxpayer Compliance Research Robert F. Boruch

339

C

Symposium on Taxpayer Compliance Research

380

D

Panel on Taxpayer Compliance Research

384

E

Committee on Research on Law Enforcement and the Administration of Justice, 1987—1988

Index

Preface This report explores what is known about individuals' compliance with federal income tax reporting requirements and how more can be learned about it. Strong concerns about this subject date back to the late 1970s, when articles in the financial and popular press publicized an "underground economy" in the United States, which purportedly accounted for hundreds of billions of dollars per year in untaxed economic activity. Legislators in search of painless but ethical ways to raise badly needed revenue quickly showed interest in this economy and sought ways to raise tax revenue from it. They asked the Internal Revenue Service (IRS) to estimate its size as well as the tax revenue lost as a result of its activity. The first IRS report (IRS, 1979b) made ingenious use of existing data to provide a ballpark estimate of the amount of economic activity that was going untaxed. Although it suggested that earlier reports had exaggerated the amount of untaxed activity, it served mainly to illustrate the limited nature of data and research on tax compliance and its causes. A second, more refined study (IRS, 1983^ placed the "tax gap" due to noncompliance with federal income tax laws by individuals and corporations at about $90 billion for 1981. When legislators and tax administrators began to search for information to assist them in developing policies to recover the revenue lost to noncompliance, they found little help in social science research. Although economists had been building theoretical models of taxpayer compliance since the early 1970s, these models were derived mainly from literature on the economics of crime and reflected too few of the unique aspects of taxpaying to provide useful policy guidance. There was also substantial work by survey researchers who obtained taxpayers' self-reports of previous noncompliance with tax laws and related them to measures of attitudes, beliefs, and sociodemographic characteristics. But the measures of compliance in these survey studies were very general, and the explanatory variables were usually related only indirecdy to policy variables that might be used to stimulate higher levels of compliance. Even without much information to guide them, legislators proceeded vigorously in their attempts to recover the revenue lost to noncompliance. They based major portions of the 1982 Tax Enforcement and Fiscal Re-

vili

Preface

sponsibility Act on the supposition—supported by some economic models but not others, some research within the IRS, and some survey research— that higher penalties would lead to fuller compliance. The Deficit Reduction Act of 1984 closed off certain tax shelters and broadened information reporting, the mechanism through which interest payments, dividends, and some other types of income are reported directly to the IRS before tax returns are filed. By that time, academic researchers outside the IRS had begun to publish articles using data collected by the IRS. Econometricians were primarily concerned with how tax rates and IRS administrative actions, particularly audits, affected compliance by individual taxpayers. Results suggested that lower tax rates and higher audit rates might indeed stimulate compliance. Economic theorists began to model some important aspects of the tax environment, such as features of the Internal Revenue Code and the interaction between the taxpayer and the taxing agency. Researchers from other disciplines sought to understand other important aspects of the process, such as appeals to morality, patriotic motives, the disapproval of peers, and imperfect knowledge of compliance requirements. This work, although best regarded as pioneering rather than definitive, illustrated for both researchers and the IRS the potential benefits to be gained from cooperation. Some might question the scholarly propriety of such cooperation. Some of the lowest compliance rates occur among small-scale entrepreneurs, people in need of income from part-time jobs, family farmers, and other generally law-abiding citizens—causing some to wonder whether noncompliance with tax laws requires the attention of researchers outside the IRS. But noncompliance with tax laws raises some serious social problems. Tax evasion by people who may have special opportunities to conceal income is unfair to others who pay their full share. Public knowledge that successful evasion occurs may encourage further evasion, leading eventually to more coercive tax collection methods. And a large underground economy may encourage attitudes of guardedness among acquaintances, dishonesty in financial transactions, and alienation from government. Taxpayer noncompliance presents intellectual as well as social problems. It offers fertile ground for the refinement and testing of theories from a host of social sciences. Such concepts as maximization of expected utility, game theory, social sanctions, guilt neutralization, social networks and social fields, decision heuristics, and bounded rationality have all been used in

Preface

ix

efforts to conceptualize the process of taxpaying. Empirical research in the area raises exceedingly difficult problems of measurement and inference— problems that challenge social scientists and produce competing claims and methodological qualifications that frustrate policy makers in search of definitive conclusions. Seeing the potential value of research but confused by often conflicting statements, the IRS Research Division in 1984 asked the National Academy of Sciences to convene a panel to assess the current state of knowledge regarding taxpayer compliance and to recommend promising areas for future research. In response, the Committee on Research on Law Enforcement and the Administration of Justice created the Panel on Taxpayer Compliance Research. The panel was a diverse group that included a former commissioner of internal revenue, leading tax lawyers and accountants, and social scientists from the major disciplines with research bases that relate to compliance—anthropology, economics, political science, criminology, psychology, and sociology (see Appendix D for brief biographies of panel and staff members). Although all panel members were respected members of their disciplines and professions, some had not previously considered tax compliance issues in depth. Fortunately, every panel member recognized that the expertise of the others was needed to encompass the diverse and complex issues related to taxpayer compliance. To fill in possible gaps and to extend the panel's knowledge, we commissioned a number of papers on specific topics; these were presented and discussed at a conference held at South Padre Island, Texas, January 15-17, 1986 (see Appendix C for the program and list of participants). Rather than simply reviewing the literature on compliance, most of the papers sought to extend knowledge by drawing insights from work being done in other areas. Eight of the papers appear as Volume 2: Social Science Perspectives. Two of the papers are concerned entirely with methodology and are included in this volume. The paper by Peter Schmidt (Appendix A) extends recent multivariate statistical work to deal with important aspects of taxpayer compliance analysis. The paper by Robert F. Boruch (Appendix B) suggests ways experimental and quasi-experimental designs can be used to inform taxpayer compliance research. The papers in Volume 2 vary widely in terms of approach and disciplinary orientation. Several are reviews of literature from subfields of several disciplines—decision psychology, social psychology, economics, and deterrence research—highlighting issues that have received too little atten-

X

Preface

tion in previous taxpayer compliance research. Others present new conceptual frameworks for compliance, and three contain original research—a microeconomic model that incorporates tax advisers, an ethnographic study of noncompliance in one occupational setting, and a history of recent federal tax and budget legislation that has defined the options available to the IRS for tax administration. Because of their disciplinary orientations and personal interests, panel members placed different levels of emphasis on the information and ideas developed at the symposium, in other prepared background materials, in meetings with compliance researchers, in extensive panel discussions, and in early drafts of this report. Some placed more emphasis on alternative theories of decision making under uncertainty. Others believed it is more important to recognize that compliance is embedded in the social, cultural, and financial sociocultural aspects of taxpayers' environments and to study how those environments develop. The tax practitioners, administrators, and policy analysts were occasionally bemused by the social scientists' arcane discussions. They regularly reminded the group of the legal, administrative, and institutional structure that defines taxpaying. "What does this debate mean for policy?" was a frequent question. Despite their differences in emphasis, panel members were in general agreement that existing economic theory reflected the institutional, legal, and administrative structure surrounding taxpaying too poorly to serve as a basis for policy. We also noted the lack of rigorous theory related to compliance from other disciplines and concurred in recognizing the need to develop more informed theories of compliance. In developing our list of important aspects of taxpayer compliance behavior, we drew on compliance research in regulatory fields. We also believe that both ethnographic and experimental research should be used to inform theory and to provide preliminary tests of models of compliance. The panel found existing empirical work flawed, but we agreed that some consistent insights regarding compliance have emerged from this work. More sophisticated empirical work on better, more current data is urgentiy needed. Improvements are urgently needed in compliance measurement at the individual level. In recommending promising areas for future research, the panel shared a hope that results would cumulate to form a firm base for tax policy and tax administration, at the same time contributing to basic knowledge in several disciplines. Some of this accumulation should occur spontaneously, as independent scholars seek to fill the knowledge gaps that are pointed out in

Preface

XI

Chapters 2 and 3. Additional important knowledge can be developed through the more applied tax administration research illustrations that appear in Chapter 4; these will require closer cooperation among researchers, tax administrators, and tax practitioners. The IRS will no doubt continue to play a central role in taxpayer compliance research, as it provides compliance research data and other forms of support. While different panel members would place different weights on various conclusions and recommendations, overall censensus was achieved. Panel members agreed that carefully designed randomized field experiments with policy innovations, when feasible, could produce the firmest conclusions about influences on compliance. Beyond that, some panel members would emphasize deductively derived theory tested by multivariate statistical analyses of audit-based data, while others would place more emphasis on exploratory and ethnographic research as a means of refining theory. As the panel's work progressed, all of us gained respect for disciplinary and administrative perspectives that were new to us. The dedicated efforts of the staff have been central to the work of the panel. Jeffrey Roth, study director from the inception of the panel, made a very substantial contribution to the panel's activities. He skillfully managed the affairs of the panel, drafted large segments of the report, guided the panel through an extensive report review process, and revised (many times) all report materials. John Scholz, whose association with the panel began shortly after its inception, was responsible for drafting significant segments of the panel report, particularly the material developed in Chapters 3 and 6. Like Jeff, John has reviewed and helped to revise various sections of the report many times. The panel also benefited considerably from the administrative and secretarial work of Gaylene Dumouchel and Teresa Williams of the National Research Council and Helen Graham of Wellesley College. Early drafts of the report were reviewed by many people. The panel and staff are grateful to members of the Commission on Behavioral and Social Sciences and Education, members of the Committee on Research on Law Enforcement and the Administration of Justice, and selected outside reviewers (Harold Grasmick, Karyl Kinsey, Robert Mason, Daniel Nagin, and Louis Wilde). Their valuable comments were all seriously considered and helped us to sharpen our thinking. Nearly all sections of the report were revised in major ways to reflect important new insights and clarifications as a result of these individuals' comments on our work. The task of editing the large volume of material assembled by the panel has been considerable. Christine McShane, editor of the Commission on

xii

Preface

Behavioral and Social Sciences and Education, not only sharpened our language but also challenged our assertions when they were insufficiently developed or documented. In doing so, she made an excellent and important contribution to the report. An important feature of the panel's work has been the support and encouragement of the sponsor, the Research Division of the Internal Revenue Service. As the 1RS contracting officer's technical representative, William Lefbom kept in close touch with the panel throughout its work and provided valuable encouragement, insights, and information. Patricia White, a member of his staff, prepared useful computer analyses and provided other information. Roscoe Egger, commissioner of internal revenue during much of the panel's life, deputy commissioner John Wedick, and assistant research director Irene Sherk also provided general encouragement and support for our efforts. Sadly, as this report was in final preparation, 1RS research director Frank Malanga passed away. Frank understood the sometimes serendipitous ways in which social science research might improve tax administration, he respected the panel's needs for independence and time to reach consensus, and he valued a broad range of research perspectives. He will be missed by the compliance research community. Ann Dryden Witte Chair, Panel on Research on Taxpayer Compliance

Summary

According to Internal Revenue Service (IRS, 1988) estimates, individuals failed to report between $70 and $79 billion in federal taxes due on legal income they received in 1986, amounting to nearly 20 percent of federal personal income taxes due and 40 percent of the federal deficit in that year. Other forms of noncompliance—for example, by organizations, by those who fail to pay the income tax they do report, and by those who receive illegal income—probably push the annual "noncompliance tax gap" above $100 billion. Since the I R S manages to recover at most an estimated 10 to 15 percent of the noncompliance tax gap through all its enforcement activities (IRS, 1986b), it has a pragmatic interest in learning why some people comply with reporting requirements, while others fail to do so. The question has also intrigued scholars from several disciplines, as well as the community of practitioners who prepare tax returns and advise clients concerning tax matters. The panel of scholars and other profesionals responsible for this report was created in response to an IRS request to the National Academy of Sciences to critically review previous research on the factors that influence taxpayer compliance with reporting requirements. The panel was asked to state the conclusions that can be drawn from this literature, to identify the gaps in knowledge that future research may be able to fill, and to suggest how the needed research might be carried out.

Defining Taxpayer Compliance Compliance with income tax reporting requirements involves accurate reporting of taxable income, accurate claims of subtractions such as income adjustments and itemized deductions, correct computation of tax liability, and timely filing of the tax return. These tasks require reading and arithmetic skills, record-keeping effort, and judgments that challenge the capabili-

2

Summary

ties of many taxpayers. Therefore, although noncompliance with tax reporting requirements sometimes results from deliberate choices, it may also occur because of carelessness, errors, and misunderstandings of requirements related to keeping records and filing returns. For this reason, the panel adopted a definition of compliance that implies no assumptions about taxpayers' motivations: Compliance with reporting requirements means that the taxpayer files all required tax returns at the proper time and that the returns accurately report tax liability in accordance with the Internal Revenue Code, regulations, and court decisions applicable at the time the return is filed.

Against this standard, noncompliance includes both overreporting and underreporting of tax liability. It includes both deliberate underreporting that is punishable by criminal sanctions and underreporting due to misinformation, misunderstanding, negligence, or some other cause. It does not include the structuring of one's financial affairs within the law so as to reduce taxes, perhaps in ways that were not intended by lawmakers. It also does not include situations lacking a clear legal precedent, in which compliance status is ambiguous.

Patterns and Trends in Taxpayer Compliance Taxpayer compliance is notoriously difficult to measure. But, according to the IRS Taxpayer Compliance Measurement Program (TCMP) for individual returns (see Chapter i, Supplementary Note B, for more explanation of this program, which involves audits of randomly sampled returns) and other IRS (1979b, 1983^ 1988) studies, the following rudimentary picture of taxpayer compliance and noncompliance patterns emerges. 1. Taxpayer noncompliance is widespread: In nearly half of all 1979 T C M P audits, the auditor recommended tax changes because of apparent noncompliance, and at least one change in an itemized deduction was recommended for 88 percent of taxpayers who itemized. 2. Unreported income accounts for about 75 percent of the amount of individual noncompliance with reporting requirements, overstated subtractions account for about 13 percent, nonfiling for about 10 percent, and arithmetic errors for the remaining 2 percent.

Summary

3

3. Not all noncompliance is in the taxpayer's favor: of taxpayers who misreport income about 12 percent overreport, of those who misreport subtractions about one-third fail to claim all to which they are entitled, among nonfilers about 40 percent have already had enough tax withheld and many would receive refunds if they filed. 4. Most noncompliance involves fairly small amounts: the 1982 T C M P cycle found that 48 percent of all returns had precisely correct reports of taxable income, 70 percent were deemed accurate within $50, and 90 percent within $1,000. As a compliance measurement device, auditors' recommendations are subject to two kinds of errors. First, T C M P audits detect less than $1 of every $3 of unreported income that is not subject to withholding or third-party information reporting (e.g., Form 1099 for interest and dividends), according to the I R S (1988). Second, they overstate noncompliance that auditors do detect by at least 12 percent, the fraction of recommendations that are reversed on appeal; some potentially successful appeals are presumably not undertaken because of the cost and uncertainty involved.

Tax Administration Strategies and Social Science Research The I R S pursues four basic strategies for administering tax laws. First, it can pursue deterrence and revenue recovery objectives either by applying resources to detect noncompliance or by designing the taxpaying process to make noncompliance easier to detect. The agency uses labor and computer resources to do such things as audit tax returns, match third-party information reports, and conduct detailed investigations. Tax administrators and legislators can make noncompliance easier to detect by such structural measures as making income more visible through information reports or by raising the minimum threshold (e.g., as a percentage of adjusted gross income) for deducting an expenditure such as medical costs. Previous research clearly demonstrates that structural measures encourage compliance and suggests that audits and computerized examinations of tax returns slightly improve compliance, both by taxpayers who are directly contacted and by those who learn about the experiences of others. But estimates vary of how responsive compliance is to strategies of deterrence and revenue recovery. Second, lawmakers and the I R S can decrease the taxpayer's cost of

4

Summary

compliance by making compliance easier: by simplifying the tax laws, by defining compliance requirements in terms of information that taxpayers may routinely record for other purposes, by simplifying tax forms and instructions, by providing informational publications, by answering telephone inquiries about compliance requirements, and by providing assistance in resolving taxpayers' problems. Research has begun to measure compliance costs, and some preliminary evidence suggests that efforts to reduce compliance costs may encourage compliance somewhat. But there are no firm estimates of how responsive compliance is to these costs or to IRS or congressional efforts to make compliance easier. Third, the IRS can encourage normative support for compliance, encouraging taxpayers to reframe their compliance decisions in a more positive light—by personalizing the process, by reminding taxpayers of social commitments, or by reminding taxpayers of the tax-supported services they receive, for example. Evidence of the compliance effects of these techniques is too sketchy at present to permit conclusions. Fourth, some IRS programs may encourage compliance less directiy through tax practitioners, who interact with many taxpayers. Over half of all returns each year are affected by lawyers, accountants, and others who prepare returns or offer advice about tax matters. These practitioners review far more returns each year than the IRS does, and they are subject to penalties for helping their clients evade taxes. Yet they are also obligated to interpret arguable positions on behalf of their clients rather than the government. Research evidence is beginning to suggest that practitioners' actions may encourage compliance with straightforward requirements and may discourage compliance with more ambiguous ones. However, the issue is far from settled. Integrating these four components into a coherent approach to tax administration is difficult, particularly since they must compete for limited resources. Furthermore, pursuit of one strategy (e.g., deterrence through fear of penalties) may inhibit the success of another (e.g., increasing commitment to comply). And some policies are two-edged swords that may encourage compliance by some taxpayers but stimulate noncompliance by others. Tax administrators are better placed than social scientists to address many of the questions that arise in the course of refining tax administration policy—for example, estimating administrative costs and assessing political constraints. But social science research offers great potential to estimate the compliance effects of policy changes, which administrators usually over-

Summary

5

look precisely because they are difficult to measure. Research can also point to characteristics of taxpayers and their environments that are beyond IRS reach but that influence compliance and should therefore be considered in developing policy. Even though taxpayer compliance research is no more than twenty years old, it has already established rather strong findings about a few factors that affect compliance. Yet knowledge is still limited because of the fragmentation of academic disciplines, the difficulties of measuring compliance and of gaining access to some of the most useful IRS data, and the difficulty of developing theory and designing empirical research that adequately capture many of the institutional realities of taxpaying. The next major section summarizes the major findings from previous research on taxpayer compliance and suggests questions that researchers should address in extending this work. The following section introduces some ideas that have proven useful in understanding other kinds of human behavior and suggests ways to apply them to taxpayer compliance.

Taxpayer Compliance Research: Findings and Prospects Research to date tends to support tax administrators' long-held conjectures that financial incentives, social sanctions, and moral commitment may all affect taxpayer compliance with reporting requirements. But much more empirical and theoretical work is needed to confirm the existence of these relationships, to measure the strength of their influences on compliance, and to learn how these factors are themselves influenced by taxpayers' environments. FINANCIAL SELF-INTEREST AND COMPLIANCE

Using both survey research and econometric techniques, many researchers have studied how compliance is affected by several elements of financial self-interest: the probability that a noncompiler will be detected, the penalties for noncompliance, the tax rate structure, and the level of income. The following sections summarize conclusions from research to date and pose questions that would be valuable for scholars to pursue in the future. DETECTION PROBABILITIES FOR NONCOMPLIERS.

Previous research

suggests that: 1. Income that is more visible to the IRS (e.g., through withholding or information reporting) is likely to be more fully reported than other income.

6

Summary 2. Experiencing an audit may slightly increase future compliance, at least if the audit discovers all unreported taxes and the taxpayer feels the outcome is fair. 3. For at least some classes of taxpayers, higher audit rates in local geographic areas may increase compliance slightly by signaling others that they are vulnerable to audits—a general deterrence or ripple effect stemming from awareness of an I R S audit presence. 4. The I R S succeeds to some extent in concentrating its audit resources on taxpayers who are least compliant.

These findings leave many questions unanswered. Many of the relevant studies are based on data that describe behavior more than fifteen years ago, and so they do not indicate how compliance has changed since then because of refinements in I R S allocation of resources and procedures for detection of noncompliance. The studies rely on audit rates as proxies for the probability that a taxpayer's noncompliance will be detected; but that probability is unknown to both the taxpayer and the I R S , and almost nothing is known about how taxpayers assess detection risks using information they have. Perhaps most important from a policy standpoint, no sound estimates are available of the magnitudes of these effects—of how responsive compliance is to enhanced information reporting, to previous audits, or to changes in audit rates. T w o extensions of existing lines of research can help to fill these gaps in knowledge. 1. Updating the 1969 data base on which most of the conclusions about general deterrence are based, if a more adequate compliance measure can be developed and validated; and 2. Measurements at multiple points in time of taxpayers' perceptions of the risk that noncompliance will be detected, and ethnographic and survey research on how those perceptions form and change in response to changes in financial circumstances, I R S contacts, and other events. OTHER E L E M E N T S OF F I N A N C I A L S E L F - I N T E R E S T .

Research to date

provides even less firm conclusions about how other elements of selfinterest affect compliance. The panel recommends that existing lines of economic and econometric research in three areas be extended to improve understanding of these relationships. 1. Noncompliance penalties. Most studies have failed to demonstrate that higher penalty rates encourage compliance. A few studies have found such an effect, at least when the probability of imposing the penalties exceeds

Summary

7

some minimum threshold level. Research is needed on the level of the threshold compared with the actual probabilities of severe penalties for noncompliers. Recent increases in prescribed penalties and fines present opportunities for before-and-after studies that could clarify how penalties affect compliance. 2. Tax rates. Some research suggests that lower tax rates may improve compliance by reducing the rewards for evasion. But other empirical research fails to confirm that effect, and there are sound reasons to doubt this finding. Since the tax structure links the rates to income levels, it is difficult to disentangle tax rate effects from income effects. Analyses of experience before and after the Tax Reform Act of 1986, which changed the tax structure, may lead to more clear-cut conclusions about how changes in tax rates affect compliance. 3. Income. Neither theoretical nor empirical findings about the relationship between true income and compliance are at all clear-cut. Achieving more understanding will require analyses of individual-level data on changes in income level and composition, experience with IRS contacts, use of tax practitioners, and compliance. Collecting and analyzing these data over a time period that includes changes in tax rates would make it easier to distinguish income effects on compliance from tax rate effects. 4. Microeconomic theory. To design future econometric studies, microeconomic models of compliance should incorporate the multiple roles of tax practitioners and recognize the different costs of compliance and risks of noncompliance detection and punishment that are associated with different return items. Models with these features can address how compliance is affected by the timing of income receipts and tax payments (e.g., through withholding, estimated tax payments, and refunds), and by various costs of compliance and noncompliance. Eventually, these theoretical models may provide a basis for empirical models that distinguish more adequately the contributions of taxpayers, taxing agencies, and tax practitioners to compliance. SOCIAL SANCTIONS AND COMPLIANCE

A few surveys of self-reported compliance suggest that social sanctions—disapproval by friends and acquaintances—may discourage noncompliance, perhaps even as effectively as legal sanctions for some taxpayers. However, because survey data reflect what respondents tell interviewers rather than actual compliance, they are not sufficient to establish this proposition. Individual-level data over time are needed that

8

Summary

link measures of the threat of social sanctions to measures of compliance constructed from tax returns or audit results. There is also a need for research on why members of different groups approve or disapprove of others' noncompliance. MORAL COMMITMENT A N D COMPLIANCE

Not surprisingly, survey research has consistently found that taxpayers who report high moral commitment to obey tax laws are unlikely to report cheating on their taxes. However, it is not clear whether this pattern reflects actual behavior or merely a desire to report behavior that is consistent with one's proclaimed attitudes. Again, survey data alone are not sufficient to establish this proposition. Rather, data are needed that link measures of social commitment to nonsurvey compliance measures. Studies of how taxpayer compliance and commitment are related to attitudes toward law generally, toward government, toward tax laws, and toward enforcement methods have produced inconsistent results. Research is needed not only on how these attitudes relate to compliance, but also on the experiences and circumstances that shape these attitudes over time. DEMOGRAPHIC PATTERNS OF COMPLIANCE

Older taxpayers appear to have higher levels of compliance. However, it is not clear whether taxpayers grow more compliant as they age, or whether successive cohorts of taxpayers are becoming less compliant. Distinguishing between aging effects and cohort effects is essential for predicting future compliance levels but requires compliance measures collected over time for individuals whose ages and other relevant characteristics are known. Compliance rates are higher for women than for men. There is also some evidence that compliance is higher among whites than nonwhites, but the differences are small and depend on details of individual studies.

Expanding the Framework of Compliance Research Previous compliance research has focused almost exclusively on the four kinds of influences just discussed: financial self-interest, social sanctions, moral commitment, and demographics. But several characteristics of taxpaying and tax administration may also affect compliance and should be more fully recognized in future research. Tax administration blends the characteristics of criminal and administrative law enforcement. Although criminal punishments are available in

Summary

9

certain circumstances, taxpayer noncompliance does not arouse social condemnation of the sort that is triggered by many street crimes, and the victim of noncompliance is diffuse and thus not salient to most taxpayers. The objective circumstances that structure taxpaying arise from two-way interactions between taxpayers and their social environments, particularly their coworkers and others with whom they carry out financial transactions. Taxpayers are only partially informed about their own objective circumstances, including the requirements and costs of compliance and the risks and rewards of noncompliance. Their methods of coping with uncertainty and misinformation may introduce erors into perceptions and decisions, which in turn contribute to noncompliance. Individual taxpayers place different moral values on compliance; these values may be influenced by such factors as childhood socialization to obeying rules, the values of friends and business acquaintances, and taxpayers' personal experiences with the tax system. These observations point to several kinds of studies that have been underused in previous taxpayer compliance research and that the panel believes should be more widely used by researchers in the future: 1. Laboratory experiments to study how tax returns are affected by the various routines that taxpayers use to cope with the complexity and uncertainty of taxpaying; 2. Ethnographic studies of individuals to learn how they form their understandings of compliance requirements, their perceptions of the financial and social risks and rewards of compliance, and the moral values they place on compliance; and 3. Ethnographic studies of social and occupational networks of individuals to learn how patterns of financial transactions develop that encourage or discourage compliance, how information (and misinformation) about compliance requirements and contacts with tax administrators are communicated, and how values concerning compliance are exchanged. These three kinds of compliance research are perhaps best carried out by individual scholars with only minimal coordination with tax authorities. In addition, the differences between tax laws and most criminal laws suggest that tax authorities adopt a broader approach to their own research. As a start, there is a need for tax administrators to examine how changes in enforcement methods affect compliance, not merely the revenue yield from enforcement. There is a need to examine how compliance responds to features of the tax system that are influenced by the tax legislators and administrators who construct it and by the lawyers, accountants, and other

io

Summary

tax practitioners who help taxpayers interpret and use it. This research will require much closer coordination between researchers, tax administrators, and the other communities. To illustrate this broader orientation to tax administration research, the panel selected three programs that could produce substantial benefits to tax administration and to the understanding of taxpayer compliance. The three programs concern: (i) responses to changes in tax laws and their administration, with special emphasis on the Tax Reform Act of 1986; (2) tax practitioners and their relationships to taxpayer compliance; and (3) the compliance effects of 1RS contacts with taxpayers. There are, of course, other excellent tax administration research opportunities, including ethnographic research in conjunction with programs such as these three. RESPONSES TO C H A N G E S IN T A X L A W S AND ADMINISTRATION

By substantially changing the tax structure, the Tax Reform Act of 1986 (TRA86) offers a remarkable opportunity for learning about taxpayers' responses to changes in the incentives that may affect compliance, to efforts to improve the perceived fairness of the tax system, and to legal changes that affect the complexity of compliance requirements. A research program on these matters should contain the following elements: 1. A comparative analysis of individual-level 1985 and 1988 TCMP data, to learn how compliance changed in response to the changes in incentives created by TRA86; 2. A survey of how taxpayers sought information about the changed compliance requirements and how TRA86 affected their use of tax practitioners' services; 3. Survey research to analyze how TRA86 affected perceived fairness of the tax system, combined with a study of how compliance changed in response to changes in perceived fairness; and 4- Organizational studies of how 1RS procedures adapted formally and informally to requirements and constraints imposed by TRA86. The panel recommends that the 1RS develop a permanent data collection and analysis capability to take advantage of similar research opportunities that future changes in tax laws and administrative procedures will present. T A X PRACTITIONERS AND COMPLIANCE

The influence of practitioners on taxpayer compliance has only recently begun to attract broad research attention, and very little is known about

Summary

n

their effects on compliance. A research program on this subject should include the following components: 1. Cross-sectional and panel surveys on the motivations, circumstances, and events that affect taxpayers' use of practitioners and their choices of particular practitioners; 2. Research on how different types of practitioners affect compliance, through analyses of individual-level T C M P data, through further organizational studies of practitioners and their interactions with clients, and if the necessary cooperation can be obtained, through analysis of the effects of an "exemplary practitioner" program administered jointly by the I R S and appropriate professional societies; and 3. Experimental studies of how specific practitioner interactions with clients (e.g., a simple probe for unreported income) affect their clients' compliance. TAXPAYER CONTACTS AND COMPLIANCE

Although taxpayer audits and the annual mailing of tax forms and instructions are perhaps the most visible ways in which the I R S contacts taxpayers, there are others. These include contacts with individuals initiated both by the I R S (e.g., computerized notices of arithmetic errors and other discrepancies, outreach classes, publications) and by taxpayers (e.g., telephoned questions, letters of inquiry, requests for problem resolution). Also included are mass media campaigns to encourage compliance through news articles and advertisements. These contacts may indeed affect compliance, but virtually no measures are available of those effects. T o measure the effects of such programs, the I R S should consider adopting a more systematic approach to evaluation and innovation. The panel recommends an approach that involves: 1. Evaluating the compliance effects of current practice using quasiexperimental designs that exploit natural variation over time and across districts in the timing, coverage, and quality of taxpayer contact programs; 2. Using laboratory experiments to design and pretest innovations in taxpayer contacts to remedy problems discovered in evaluations; and 3. Testing promising innovations in contacts using randomized experimental designs in the field wherever possible, to measure compliance effects and revenue yields more accurately. Various psychological theories suggest some specific tactics for improving the effectiveness of I R S communications. These include personalizing

12

Summary

letters by providing a named IRS staff member, reminding taxpayers of instructions that may apply to their particular situations and that may reduce their tax liabilities, and calling attention to compliance motivations that may appeal to taxpayers as good citizens. Besides potentially improving tax administration, laboratory and field experiments involving techniques such as these would provide further tests of the relevant psychological theories.

Data Needs in Taxpayer Compliance Research Much taxpayer compliance research requires data on large samples of taxpayers, including measures of compliance and of other variables that may help to explain compliance. Suitable data bases have been produced in four ways: 1. The T C M P Survey of Individual Returns Filed, an IRS program conducted every three years, which involves comprehensive audits of random samples of about 50,000 tax returns in each cycle; 2. Taxpayer surveys that collect self-reports of previous compliance and data on relevant taxpayer characteristics and attitudes; 3. Experimentally generated data bases, in which the compliance effects of a randomized treatment are measured as the average difference between the treatment and experimental groups in year-to-year changes in tax return items; and 4. A 1969 aggregate cross-sectional file (IRS Project 778) that contains, for each three-digit zip code area, an estimate of the aggregate compliance level, data on IRS activity, and demographic and socioeconomic characteristics of the area. Each kind of data has particular strengths but fails in some way to adequately support research on the determinants of taxpayer compliance. To enhance the value of these four data collection approaches, the panel makes six recommendations. 1. To expand the research uses of TCMP, we recommend that the IRS consider modifying the TCMP check sheet to include a more detailed description of the preparer who signs the return, information on others whose advice may have affected the return, information on the taxpayer's primary and secondary occupation and industry, and information on the issues underlying changes the auditor recommends on the return.

Summary

13

2. To improve the timeliness of TCMP data for all its uses, we recommend that the IRS formally consider restructuring the program from a three-year cycle to an annual or a biennial cycle (with a fraction of the entire sample audited in each year) and search for ways to reduce the three- to four-year delay in file development. 3. To enhance the value oftaxpayer surveys for reseachers and tax administrators, we recommend that the IRS and others who undertake surveys expand their research focus—to include the relationships between values, perceptions, and compliance; to give greater emphasis to the processes, circumstances, and events that affect the relevant values and perceptions over time; and to provide more resources for the replication of findings. 4. To strengthen the foundations of survey research on taxpayer compliance, we recommend a research program to improve the interpretation of self-reports of noncompliance and to ascertain whether and how the specificity and validity of self-reports can be improved through the randomized response technique, innovations in questionnaire design, and repeated interviews. 5. As a possible basis for more current information about the general deterrence effects of IRS activity, we recommend that the 1969 aggregate cross-sectional data base (Project 778) be updated and refined if the reliability of the compliance measure can be demonstrated for a more recent year. 6. To provide a basisfor stronger inferences about the compliance effects of administrative activities, we recommend that the IRS and external researchers increase their use of randomized experiments in both field and laboratory settings. Taxpayer compliance research, including several projects recommended in this report, raises very difficult questions about how to provide useful knowledge without jeopardizing taxpayer privacy, other generally accepted rights of research subjects, and the IRS mission. To deal with these matters, the panel recommends long-term cooperative efforts—involving the IRS and the broad research community—to deal with three issues: 1. To develop and implement policies governing access by researchers to compliance research data, especially external use of sensitive IRS data. 2. To develop and implement policies to foster compliance research within the IRS that uses the most methodologically sound research techniques, including randomized experiments when needed, to ad-

14

Summary

dress problems in tax administration while protecting research subjects from unfair burdens and risks. 3. To formally and thoroughly consider the feasibility and desirability of arranging for the deposit of sensitive IRS data in an independent secure repository. A standing authority should be established to interpret and monitor these policies and arrangements in light of evolving research ideas, changes in compliance requirements and forms of noncompliance, innovations in tax administration, and advances in statistical matching techniques. Additional resources should also be provided for the IRS Research Division to continue and expand its technical effort to support external researchers' use of IRS data bases through public-use files or special tabulations.

A Look Ahead Much has been learned in twenty years of taxpayer compliance research— but much remains to be learned that is of intellectual and policy interest. The understanding of taxpayer compliance can best be advanced through the involvement of an intellectually exciting community of researchers that extends across the tax professions and social sciences. While such a community has already begun to form, its growth can best be encouraged by a diverse array of organizations—including private foundations, federal research sponsors, and the tax policy community, as well as the Internal Revenue Service. The IRS can continue its major role in compliance research by improving the research potential of its data bases, facilitating expanded research use of its data, integrating research and evaluation more fully into its policy deliberations, and strengthening the infrastructure that supports the research community. To assist the IRS in these efforts and to provide a structure for sustained communication with the research community, we recommend that it form and regularly convene a Compliance Research Advisory Group that brings external experts in taxpayer compliance research together with agency administrators to regularly review research priorities and revise them as conditions change. Consideration should also be given to the possible need for new organizational arrangements to strengthen and broaden the compliance research community. These arrangements might include a center

Summary

15

for compliance research, a consortium of sponsors supporting a range of activities, or a looser federation of cooperating institutions. It is essential, of course, that any such organization not divert significant resources from original research, and that it serve to complement and facilitate the work of individual researchers, existing tax research centers, and established data centers.

i. Paying Taxes

This report is concerned with why people comply with federal income tax reporting requirements—a question of both intellectual and policy interest. Policy interest has grown dramatically in the last decade, in an environment of large budget deficits, reluctance to raise taxes, and the discovery of a large "underground economy" and other forms of noncompliance. The Internal Revenue Service estimates that for 1986 the "tax gap" arising from individuals' failure to report federal income taxes due was $70 billion— about 40 percent of the federal deficit (IRS, 1988). Noncompliance with other federal income tax requirements—by corporations, by those who receive illegal income, and by those who fail to remit taxes due, for example—probably pushed the total tax gap above $100 billion. Noncompliance of this magnitude raises important issues of equity between citizens who pay their taxes fully and those who do not. For some, it also raises questions about the future viability of the primary system for supporting federal services. In recent years the IRS and other institutions in the tax policy community—other units within the Department of the Treasury, the tax committees of Congress, the General Accounting Office (GAO), the Tax Section of the American Bar Association (ABA), and the American Institute of Certified Public Accountants (AICPA)—have all undertaken projects concerned with improving compliance. At least one 1988 presidential candidate highlighted improved compliance as a means of reducing the federal deficit. This rising tide of concern contributed to the passage of statutes in 1981, 1982, and 1984 that were intended to strengthen tax law enforcement and culminated in the Tax Reform Act of 1986. These legislative initiatives sought to improve compliance through a variety of tools: new and increased penalties for various forms of noncompliance, expanded requirements for information reporting to the IRS (e.g., reporting of interest and dividends by banks and brokerage firms), new restrictions on tax shelters, and changes in compliance requirements that removed some

Paying Taxes

17

taxpayers from the tax rolls and reduced the number who could benefit from itemizing deductions. T o improve its ability to cope with noncompliance, the I R S has initiated administrative innovations on a variety of fronts: the use of computers to compare reports of transactions on tax returns with information reports, simplification of forms, and refinement of the statistical algorithms used to select returns for audit. The implicit rationale for this legislative and administrative activity seems to be that simplicity, equity, and fairness of the tax code, combined with efficient but unobtrusive enforcement,' are keys to improving taxpayer compliance (see Dubin, Graetz, and Wilde, 1987c; Steuerle, 1986; and Scholz, Vol. 2, for more complete discussions of trends in tax administration). For the scholar, understanding the broad range of behaviors involved in taxpayer compliance requires the intellectual tools of many social science disciplines and provides ample empirical material for applying the most recent scientific advances and for challenging current paradigms. Researchers in different disciplines have attempted to understand taxpaying in different ways. Sociologists and anthropologists have studied questions about compliance effects of social status, communication of experiences and values through social networks, the cultural environment of taxpayers, and the structure of financial transactions. Economists understand taxpayers' compliance behavior to reflect a calculus comparing the expected benefits of evasion with the expected costs, including the possibility of being punished. Social psychologists focus on social norms in trying to account for changes in compliance. Paying taxes imposes complex decisionmaking demands on taxpayers, and psychologists have raised questions about systematic biases in taxpaying arising from the decision-making shortcuts that individuals use in uncertain, complex situations. This array of influences on taxpayer compliance illustrates both the wealth of topics that researchers from many disciplines could investigate and the potential value of applying knowledge from all these disciplines to tax policy and administration problems. Current interest in taxpayer compliance, especially individuals' compliance with federal income tax laws, has already stimulated a large number of academic studies, several multidisciplinary literature reviews, and some broader research programs. The I R S , the A B A , the Arthur Young Foundation, and other organizations have indicated their commitments to encourage compliance research by sponsoring research projects, by organizing seminars that bring together the academic and tax policy communities, and by facilitating the use of sensitive or unwieldy data sets. Recognizing the richness of taxpaying as a domain

i8

Paying Taxes

for building and testing theories of human behavior, the National Science Foundation has begun sponsoring independent research on the topic. Until recently, research on taxpayer compliance has been hampered by unduly narrow disciplinary and methodological boundaries, by researchers' ignorance about the institutions of taxpaying and tax administration, and by restrictions on access to the rich data resources within the IRS. Consequently the body of compliance research has only recently begun to grow rapidly, to integrate insights drawn from practical experience and diverse theories and methodologies, and to address the questions of tax administrators with some understanding of their current problems and response options. As interest in the issue grows, the time is opportune for tax administrators and policy makers, researchers, and research sponsors to discuss what is already known and what future research directions are likely to produce further knowledge and insights that may help to improve tax administration. This report is intended to contribute to that discussion.

Scope of the Panel's Inquiry The Panel on Taxpayer Compliance Research was convened in response to an IRS request to the National Academy of Sciences to synthesize existing knowledge about influences on taxpayer compliance and to recommend future research that offers the prospect of augmenting that knowledge base. The scope of the panel's report is, of course, shaped by both the nature of the request and the state of existing research. Although taxpayers and noncompliance come in many varieties, published research and working papers have examined only some of them. Published hypotheses about taxpayer compliance are nearly all grounded in theories of the behavior of individuals—as maximizers of their own selfinterest under uncertainty, or members of social groups, or actors in an environment that structures their financial transactions—rather than corporations, partnerships, and other enterprises. When compliance researchers make any explicit assumption about the tax base, they usually consider income taxes rather than sales, excise, personal property, payroll, or other taxes. Of the three major noncompliance categories—failing to file the return, failing to report tax liability accurately on the return, and failing to remit taxes owed—most theoretical work has focused on the reporting of tax liability. To the panel's knowledge, no models consider explicitly how compliance with income tax laws is affected by the legality or illegality of the income being taxed.

Paying Taxes

19

In published empirical research, all taxpayer surveys known to the panel concern individual rather than organizational taxpayers, and most of them differentiate between nonfiling, inaccurate reporting, and nonremittance. But, with only a few exceptions, available nonsurvey data permit analyses of compliance only with requirements for accurate reporting of tax liability on filed returns, not with filing or remittance requirements. With a few exceptions that deal with state taxes (e.g., Groves, 1958; Mason and Calvin, 1978, 1984; Klepper and Nagin, 1987b), most empirical research on taxpayer compliance is concerned with the federal income tax. In short, most existing taxpayer compliance research is concerned with accurate reporting of tax liability on legally earned income by individual filers of federal income tax returns. The panel therefore focused primarily but not exclusively on that subject. It would of course be of interest to develop theoretical models of return filing and tax remittance and to generalize the models of reporting described there to other tax structures and to organizational taxpayers. In virtually all cases, it is an open and interesting question whether the findings we report apply to compliance in these other contexts. But in developing explicit research recommendations, the panel concentrated on extending existing lines of research and studying the effects of other related but understudied influences—changes in compliance requirements, the effects of tax practitioners, and the psychological framing of compliance requirements—on individual reporting of federal income tax liability. This choice reflects the interests and expertise of the panel members, as well as the relative magnitudes involved. Because of differences in the taxable income involved, the tax gap due to individual underreporting of tax liability on legal income is estimated to be about seven times the gap associated with illegal income, ten times the tax remittance gap, eleven times the corporate noncompliance gap, and twenty-three times the gap associated with individuals' failure to file returns (IRS, i983f). Nevertheless, the panel strongly supports theoretical and empirical research on taxpayer compliance in these other contexts. In reviewing and synthesizing previous research, the panel attempted to consider all theoretical and empirical research on taxpayer compliance that had been published in the United States by the end of 1987. It also commissioned a number of papers on special topics related to taxpayer compliance: two methodological papers are included as appendices to this volume, and eight others appear in a separate volume. In addition, the panel considered a number of other documents that came to its attention, including unpublished working papers and conference presentations on

20

Paying Taxes

taxpayer compliance, pertinent published literature from other fields, internal IRS memoranda, and foreign research literature on taxpayer compliance. While our reviews of these other categories are by no means comprehensive, pertinent items from them are cited throughout this report.

Defining Compliance Accurate reporting of tax liability can be treated as having four components: (i) accurate reporting of income subject to tax; (2) accurate claims of subtractions such as income adjustments, itemized deductions, exemptions, and tax credits; (3) timely filing of tax returns; and (4) correct computation of tax liability. Thus, unlike compliance with many criminal laws, which require only that individuals refrain from certain proscribed activities, compliance with reporting requirements requires a series of actions that may involve substantial effort, reading and computational skill, and judgment. For this reason, noncompliance may occur in a variety of ways and for a variety of reasons other than a deliberate decision to understate tax liablity. This variety complicates the problem of defining noncompliance in either behavioral or legal terms. In Volume 2, Kidder and McEwen define a typology of noncompliance that attempts to illuminate the different forms of behavior involved. Their definitional scheme includes such terms as procedural noncompliance (e.g., a failure to observe some detail of filing rules that may or may not affect the accuracy of the report) and lazy evasion (e.g., an underreport of tip income that may be deliberate but is motivated more by the effort required to record taxable income accurately than by a calculated intention to evade taxes). The Internal Revenue Code defines a legal taxonomy of noncompliance by distinguishing between evasion—the criminal offense of willfully attempting to escape tax liability (26 U.S.C. 7201)—and negligence—failure to make a reasonable effort to comply (26 U.S.C. 6653).2 But, as noted by the American Bar Association's Commission on Taxpayer Compliance (American Bar Association, 1987) and by Smith (1986), the boundaries are fuzzy between evasion, negligence, and legal but controversial positions that reduce tax liability. The fuzziness arises because even neutral, fully informed experts sometimes disagree about what the compliance requirements are in a particular situation. This potential for disagreement is reflected in the use of such terms as substantial authority and reasonable basis in regulations that are intended to define the boundaries.

Paying Taxes

21

Thus, both behavioral and legal definitions of noncompliance invoke assumptions about taxpayers' motivations. Discovering the nature of these motivations is precisely the task that compliance researchers undertake, and it is extraordinarily difficult. Consequently, the panel found it more straightforward to adopt a compliance definition that minimizes the role of required assumptions and to define noncompliance as any departure from that standard, regardless of the explanation. Therefore, in this report: Compliance with reporting requirements means that the taxpayer files all required tax returns at the proper time and that the returns accurately report tax liability in accordance with the Internal Revenue Code, regulations, and court decisions applicable at the time the return is filed.

When the taxpayer's return reports a tax liability less than the accurate amount, we use the term underreporting. Similarly, we use the term overreporting when the taxpayer reports a liability greater than required. Underreporting and overreporting are both forms of noncompliance, as the panel uses the term.3 Under this definition, misclassifications, in which the taxpayer uses the wrong return item to report a transaction but the error does not affect the reported tax liability, are treated as compliant rather than noncompliant. Choice of this definition does not deny the broad range of phenomena that may lead a taxpayer to report income tax liability incorrecdy. For some taxpayers, failure to comply results from a willful act of tax evasion. For others, noncompliance may result from ignorance of tax laws or from simple mistakes in following instructions on the tax forms or in calculating taxes. Between these two extremes, noncompliance may be due to misinformation or misunderstandings about the tax laws. Taxpayers obtain misinformation from such sources as friends, acquaintances with whom they deal financially, tax preparers or advisers, and IRS staff who respond to telephone inquiries. Misunderstandings may reflect inattention, lack of interest, or acceptance of conventional wisdom about compliance requirements within a taxpayer's network of acquaintances; they may persist despite a sincere effort to find the appropriate interpretation of a confusing issue. Some misunderstanding undoubtedly reflects a willingness, sometimes to the point of negligence, to assert aggressive interpretations of the tax law that may reduce the taxpayer's tax liability, whether or not they are based on a reliable authority. Thus, willfulness, aggressiveness, carelessness, ignorance, good-faith reliance on misinformation, and acceptance of acquaintances' judgments are all of interest as possible explanations of failures

22

Paying Taxes

to comply—but we have excluded these judgments about the intentions and psychic state of the taxpayer from the definition of compliance. Although this definition of compliance avoids the difficulties associated with inferences about taxpayers' states of mind, it cannot escape the problem of ambiguity, which may arise when neither statutory nor case law has anticipated a particular set of circumstances. In that situation, the law does not precisely define compliance requirements, and even fully informed neutral experts may disagree about how the law should be applied. When it is necessary to designate taxpayers' reports that fall within this range of disagreement, we refer to them as ambiguous. For the most part, however, the panel treats ambiguous reports as compliant reports. Ambiguous cases create a fuzzy boundary between (noncompliant) underreporting and (compliant) reduction of tax liability. The term tax reduction is used to mean compliant behavior that reduces one's tax in a way that may be unintended by tax legislators but is permissible under the statute. It is accomplished by structuring transactions so as to minimize tax liability within the limits of the law and is legally permissible even in cases that may be morally objectionable to some people (for example, when wealthy individuals carry out advantageous transactions that produce no tax liability). Aggressive tax reduction may rely on currently unestablished legal interpretations that vary in their plausibility and possibility of being accepted by the courts. Beyond the limits of ambiguity, favorable interpretations that are asserted by taxpayers without a legal basis constitute noncompliance. In filing noncompliant reports, taxpayers may simply neglect to ascertain requirements. Or they may deliberately assert specious legal arguments or assert that ambiguity exists to disguise intent. If the underreporting is legally proven to be intentional, it constitutes tax evasion, a criminal offense. Although the panel did not treat ambiguous reports as noncompliant, there are several ways in which ambiguity may aggravate noncompliance. Exploitation of ambiguity and assertions of ambiguity by high-income, well-informed taxpayers may stimulate resentment and noncompliance by other taxpayers. The existence of ambiguity may encourage some taxpayers to use specious legal arguments in an attempt to claim that an unsupportable interpretation of compliance requirements lies within the range of ambiguity. Legislators or tax administrators, in attempting to reduce ambiguity, may increase the complexity of legislation or regulations through requirements that increase the volume of record keeping, the burden of computations, the difficulty of interpreting regulations, or the difficulty of

Paying Taxes

23

imputing values in noncash transactions. The increased complexity may aggravate noncompliance. One indicator that complexity may affect compliance is the fact that substantial fractions of noncompliers overreport rather than underreport their tax liabilities. It seems likely that at least some of the overreporting is due to errors that result from complexity. Even this brief explanation of the compliance definition adopted by the panel has hinted at the broad array of attributes—of taxpayers, of their financial and social environments, and of the federal income tax and its administration—that may affect compliance. Distinguishing knowledge about those relationships from suppositions was the panel's task. Before proceeding to that discussion, however, it is useful to describe what little is known about the extent of compliance and to summarize the administrative techniques that the I R S uses to administer the tax laws.

The Extent of Compliance The extent of taxpayer compliance is known very imperfectly. Because noncompliance cannot be directly observed, its magnitude must be inferred, usually through analyses of taxpayer audit results, discrepancies in official economic statistics, and specially collected data, such as surveys of tipping practices. The necessary data are expensive to collect, and assumptions are needed to derive the compliance estimates. Because of these difficulties, estimates are prepared infrequently and are also imprecise and controversial. The panel's primary concern was with understanding compliance rather than measuring it. Therefore, we did not critically review the available estimates, and our recommendations in Chapter 5 concerning compliance measurement address the quality of individual-level compliance measurement, for research purposes, not the accuracy of measures of the aggregate tax gap due to noncompliance. Nevertheless, while imprecision and controversy surround the aggregate estimates, it is useful to summarize them in this section as a context for our report. A more complete discussion appears in Supplementary Note A to this chapter. Taxpayer audits are a basic source for many compliance estimates, especially those that attempt to describe noncompliance patterns rather than merely totals. Under its Taxpayer Compliance Measurement Program (TCMP), the I R S regularly audits samples of taxpayers and uses discrepancies between taxpayers' reports and the T C M P auditors' assessments as a measure of noncompliance. This measure is used both in developing sta-

24

Paying Taxes

tistical rules for selecting tax returns for audit and in producing aggregate noncompliance estimates. The TCMP program is described more fully in Supplementary Note B to this chapter. Even though TCMP audits are conducted with special care and are generally acknowledged to provide the most accurate estimates available, auditor-taxpayer discrepancies—or recommended changes—measure noncompliance only imperfectly. Auditors fail to detect an unknown fraction of unreported tax liability, especially tax due on casual income that is not reported to the IRS by the income sources themselves through withholding and information reporting, such as on forms W-2 and 1099. For tax year 1976, the IRS (1983^ 1988) has estimated that for every $1 of unreported casual income that it detected, more than $3 of casual income was actually unreported. But there is controversy over that estimate, and the fraction should have decreased since then because of administrative improvements. Auditors also sometimes overassess tax liability, through error or because they interpret an apparently ambiguous report differently from the taxpayer. According to the IRS (1988), about 12 percent of the tax assessments by auditors are successfully challenged on appeal. But that correction misses auditors' overassessments that are not appealed because of the associated cost and uncertainty. Interest outside the government in measuring noncompliance first became apparent during the 1970s, when estimates appeared of an underground economy that involved several hundred billion dollars of unreported economic activity annually (Feige, 1979, 1980; Gutmann, 1977; see Henry, 1983, for a detailed comparative history of aggregate compliance measurement). Even if these estimates are accurate, they would dramatically overstate unreported taxable personal income because they fail to subtract legitimate deductions. Most of the estimates of unreported taxable personal income for 1976, one commonly used reference year, cluster between $75 and $100 billion (see Supplementary Note A, Table 1). However, the IRS (i983f) estimate was $118 billion for that year, and others offer even higher estimates. For this report, the more pertinent concept is the individual income tax reporting gap, that is, the federal income tax liability that is not reported on timely filed tax returns. The IRS (1979b, 1983^ 1988) has published three sets of estimates of this figure, using successively more refined data and methods, that have led to lower estimates. The 1988 estimates were $26.3 billion for 1976, $46.3 billion for 1982, and $70.1 billion for 1986, after appeals (successful appeals reduced the estimate based on auditors' assess-

Paying Taxes

25

ments by about 12 percent). In each year about 90 percent of the gap was attributable to discrepancies on filed returns and about 10 percent to nonfilers. These estimates are somewhat lower than earlier (IRS, 1983^ estimates, which are reported in Supplementary Note A, Table 2. While none of the figures should be considered very precise, they suggest that at least 80 percent of estimated true tax liability is being reported by taxpayers on timely returns. Nevertheless, the tax gap due to noncompliance involves substantial sums. For understanding compliance and developing policies to improve it, it is useful to know how widespread noncompliance is. One such measure is the prevalence of noncompliers—the fraction of taxpayers who fail to comply—according to the panel's definition. Prevalence has been estimated based on TCMP audit results and on surveys in which taxpayers are asked to report their own noncompliance in previous years. On the basis of 1979 TCMP audits, noncompliance is believed to be quite widespread (Supplementary Note A, Table 3). Nearly half of all TCMP audits found discrepancies or changes in taxpayers' reports of income items, with underreports exceeding overreports by a factor of 7 to 1. Fewer returns had changes in subtraction items—only 7 percent of all returns had changes in adjustment items, such as employee business expenses, and 25 percent in itemized deductions. But those rates do not reflect the fact that income items are reported on virtually all returns, while adjustments or itemized deductions are used on only some returns. As a fraction of only the returns using these items, the noncomplier prevalence rates are over 50 percent for adjustment items, and 88 percent for deductions. For subtraction items, changes that lead to underreports of taxable income exceed overreports by a factor of only 2 or 3 to i. 4 In addition, for 1979 fewer than 6 percent of all households failed to file required returns according to the IRS (Supplementary Note A, Table 3). Of these, nearly one-third would have either received a refund or had no balance due if they had filed as required. In short, all three elements of noncompliance—involving income items, subtraction items, and filing itself—sometimes injure the taxpayer rather than the government, suggesting that errors and oversights may help to account for the behavior. According to taxpayer surveys, only 12-22 percent of respondents acknowledge ever underreporting income, and 5-25 percent acknowledge "overstating deductions," the usual survey wording for reducing tax liability by overreporting subtraction items (Supplementary Note A, Table 5). As one might expect, these are low compared with estimates based on

26

Paying Taxes

audits—especially considering that the audits relate to only one tax year as a reference period, while the surveys use much longer reference periods, usually "ever" or "within the past five years." The lower values are to be expected, of course, because respondents acknowledge only instances of noncompliance that they recognize, recall at the time of the interview, and are willing to reveal to an interviewer. Most surveys that use special techniques (e.g., randomized response and locked-box approaches) to reduce respondents' sensitivity to the question by preserving their anonymity report higher prevalence estimates than other surveys. Compliance measurement through audits permits analyses that are more detailed in terms of both dollar amounts and the return items on which noncompliance occurs. For example, Supplementary Note A presents tabulations of the 1982 TCMP data performed for the panel, demonstrating that most cases of noncompliance concern fairly small dollar amounts (Table 6). While only 48 percent reported the total of their income items exactly correctly according to TCMP auditors, 70 percent reported within $50 of the auditor's assessment and 90 percent reported accurately within $1,000. This pattern is even more striking with respect to deductions: while only 8 percent of taxpayers who itemized deductions calculated them precisely according to the auditor, 52 percent were within $50 and 85 percent were within $1,000. Those tabulations also make clear that compliance depends on the structure of tax return items. The visibility of income to the IRS through withholding and information reporting has a substantial effect. While 95 percent of all taxpayers correctly reported their wages and salaries subject to withholding, only 70 percent reported all dividends (which are subject only to information reporting), and only 47 percent reported all tips. Easy access to records also seems to encourage compliance, which is better for mortgage interest paid to financial institutions (which usually provide annual summary reports) than for mortgage interest paid to individuals (who frequently do not). Possible effects of ambiguity and complexity on compliance are also suggested by statistics on the underreporting and overreporting of particular return items (Supplementary Note A, Table 6). Compliance is less prevalent for return items that involve burdensome record keeping and interpretations of complex rules than for other items that are similar in terms of visibility to the IRS and access to records. For example, underreporting is somewhat more common for Schedule E income (rent, trusts, royalties, other) than for capital gains, but so is overreporting. These and

Paying Taxes

27

other return item patterns are generally consistent with those that appear for the voluntary reporting percentage (VRP)—an estimate of the fraction of dollars being accurately reported rather than the fraction of filers who are accurately reporting (Table 6). Additional compliance patterns are described in Supplementary Note A.

Tax Administration and Taxpayer Compliance In defining compliance and describing its magnitude in the preceding two sections, it was useful to introduce concepts that have meaning only in the context of a system of tax administration. These concepts include the visibility of income to the 1 R S and the complexity and ambiguity of compliance requirements. The need for such concepts illustrates that taxpayer compliance cannot be understood without being related to the system of tax administration. As factors that help to define compliance, they also draw attention to tools of tax administration, such as third-part)' information reporting of income and simplification of compliance requirements, which can be expected to affect levels of taxpayer compliance. Because the tax administration system is so important in understanding and influencing taxpayer compliance, it is useful to describe that system in more detail before proceeding further. 1 R S activities could be classified in a number of ways. Instead of grouping administrative programs by organizational units, in this discussion we group them by compliance strategies: (1) increasing the probability that noncompliance will be detected, either by changing the tax structure so as to make noncompliance more visible or by increasing the resources devoted to detecting and penalizing noncompliance; (2) decreasing the costs of compliance; (3) encouraging compliance through public communications; and (4) regulating tax practitioners. The section concludes with brief discussions of the activities of the 1 R S Research Division and the role of research in developing policies for increasing compliance. I N C R E A S I N G T H E P R O B A B I L I T Y OF D E T E C T I O N

One way to encourage compliance is to increase the probability that noncompliance will be detected and penalized. This can be done in two ways: by structuring taxpaying so as to make noncompliance more visible to authorities and by devoting more resources to detecting noncompliance. INCREASING THE VISIBILITY OF NONCOMPLIANCE.

I f all t h e i n f o r m a t i o n

needed to correctly calculate an individual's taxes were readily available to

28

Paying Taxes

the IRS, noncompliers would face an enormous risk of getting caught. Indeed, the availability of such information would obviate the need for taxpayers to file tax returns, and, in such a formless society, compliance with reporting requirements would be automatic. Although society is far from formless for all taxpayers, the system of reporting that has been developed since withholding requirements on wages were enacted in 1942 has brought us close to that goal for the nearly 20 percent of all taxpayers who file short forms (IRS, 1987). The income tax due on wages is subject to withholding by employers. Information reports are filed by third parties for such income as interest and dividends from stocks and bonds, state income tax refunds, miscellaneous income such as employer-paid moving expenses, fellowships and royalties, and payments to independent contractors involving more than $500. Information reports are also required for certain subtraction items, such as the adjustment for contributions to individual retirement accounts (IRAs). In the current system of tax filing, a computerized Information Reporting Program (IRP) uses the third-party information reports to verify corresponding items on taxpayers' returns. The IRS received over 900 million information reports in its 1987 information returns program. The 1987 IRP generated 2.2 million notices of potential discrepancy between information documents and taxpayers' returns, and another 2.5 million taxpayers were sent notices of apparent failure to file tax returns. These figures reflect the importance of the programs but overstate the extent of noncompliance they discover, since some of the discrepancies that trigger notices are due to errors on information returns or to taxpayers' misclassifications that do not affect tax liability. In the narrowest sense, IRP is clearly cost effective, since the cost of mailing the computerized notices is low compared with the expected revenue returned in response to them. Through general deterrence, IRP may also encourage compliance by other, unnotified taxpayers who hear about IRP notices from acquaintances who receive them. However, there are possible secondary consequences that make the outcome less clear. Taxpayers who receive erroneous notices may become resentful and therefore make less effort to comply in the future. Also, some taxpayers will ignore their notices or will dispute the alleged discrepancies, thereby confronting the IRS with the need to decide whether to begin more extensive follow-up procedures. Failure to follow up may cause taxpayers to lower their estimates of the probability that noncompliance will be pursued and penalized. Their revisions may encourage future noncompliance, by both them and their acquaintances who may hear about the experience.

Paying Taxes

29

For these reasons, an IRS follow-up decision that is based solely on a comparison between cost and expected direct revenue yield will ignore large and potentially counterproductive effects on future compliance. But even though estimates of these compliance effects could be obtained through fairly straightforward applications of standard methods, the necessary research has not yet been undertaken. The effectiveness of IRP depends on the quality of information reported to the IRS. Thus, compliance with information-reporting requirements by income sources such as banks is an important part of the program. The 1981 and 1982 tax acts increased IRS authority to gain compliance with information-reporting requirements. In 1984, a Payer Master File was established to track information reporters; it identified 50,000 who apparendy stopped reporting some type of information. Better understanding of this group might be valuable in encouraging compliance efficiendy: there are fewer information reporters than taxpayers, they frequently have incentives to keep track of the requested information as the basis for claiming a business expense, and, for many of them, payments to others leave more visible audit trails than most taxpayer transactions. These possibilities are considered by Kagan (Vol. 2). The visibility of noncompliance can also be increased in other ways. The 1982 Tax Enforcement and Fiscal Responsibility Act (TEFRA) increased the visibility of tip income, a known area of low compliance, by requiring that restaurant owners report 8 percent of their gross receipts as a proxy for the actual tip income received by their employees. Tax forms now require that taxpayers seeking deductions for alimony payments and interest payments to individuals must report those individuals' identities; the receivers' returns can be checked to ensure that payments were reported. Computer match programs also keep track of items that are limited over taxpayers' lifetimes, such as the one-time exclusion of gain on a residence or the maximum residential energy credit for home improvements. Other less direct strategies to increase the visibility of transactions, such as discouraging the use of cash transactions, may be feasible but are not being seriously considered for tax administration purposes at this time (see Feffer et al., 1983; Kagan, Vol. 2). The strategy of making noncompliance more visible is pursued somewhat differendy in the case of subtractions from income, such as adjustments for professional expenses or deductions of medical expenses. Unlike some income-producing transactions, these transactions are always visible to the IRS because they are reported on the tax return. But verifying their compliance status requires revenue agents to examine the taxpayer's

30

Paying Taxes

documentation. Because these examinations are costly to the IRS and to taxpayers, they are frequently not undertaken unless a large subtraction arouses concern. Therefore, the compliance status of many small transactions remains invisible to the IRS. By raising the fraction of income that items such as medical expenditures must exceed before they can be subtracted, the Tax Reform Act of 1986 made it cost effective for the IRS to examine a larger fraction of these subtractions. Therefore, their compliance status became more visible, on average, to the IRS. INCREASING RESOURCES FOR NONCOMPLIANCE DETECTION. The most cosdy and familiar enforcement function of the IRS is the examination, or audit, of tax returns. More than one-third of the IRS budget is used for this function, which involves over 16,000 revenue agents who handle complex corporate and individual returns and 3,100 tax auditors who conduct taxpayer audits in IRS offices (IRS, 1987). The percentage of returns audited has dropped steadily over the last decade. In 1987, about 1,100,000 individual returns were audited, only 1.09 percent of the individual returns filed. This represents a drop from 2.3 percent in 1975 (see IRS, 1975). High-income returns are audited at higher rates: 1.4 percent of taxpayers with total positive income in the $25,000-50,000 range, 2.2 percent of those with income over $50,000, and nearly 4 percent of those with income over $100,000 who file Schedule C were audited in 1987. In addition, a statistical Discriminant Index Function (DIF) is used to score all returns with an indicator of the potential yield from audit (see Supplementary Note B for further explanation of the DIF system). Although the fraction of returns examined has decreased in recent years, use of the DIF system has clearly improved the IRS's ability to select the returns for which audits are most likely to result in tax changes. Together with improved auditing techniques, the improved selectivity has also increased the average revenue yielded per audit. In 1975, 81 percent of revenue agent and 76 percent of tax auditor audits resulted in tax changes; audits by revenue agents yielded an average of $2,584 in recommended tax change and penalty, and audits by tax auditors yielded an average of $220. In 1987, 88 percent of revenue agent and 86 percent of tax auditor audits resulted in tax changes, with an average tax change and penalty of $5,922 for revenue agents and $1,020 for tax auditors (IRS, 1987, Table 8, adjusted to 1975 dollars). But the revenue yield from audits is only part of the story. Their yield amounts to only a small fraction of the noncompliance tax gap, and there are limits to the extent that this fraction can be increased. As explained more

Paying Taxes

31

fully in Supplementary Note A, the IRS (1986b, Table 2) estimates that audits and IRP matches of individual taxpayers together yielded $3.2 billion in additional taxes and penalties on individual returns for 1985— about 5 percent of the amount underreported by return filers. The same report also estimates that only about $28 billion—about 31 percent of the individual noncompliance tax gap including nonremittance—could be recovered through all IRS enforcement programs combined, even if they were increased to the extent possible without incurring marginal costs in excess of their marginal revenue yield. As explained by John Scholz (Vol. 2), the IRS has found it impossible to augment its budget sufficiently to achieve this economically feasible limit. The limited ability of the IRS to recover revenue lost to noncompliance suggests why it should be concerned with maintaining and increasing the compliance that occurs without agency follow-up—"voluntary compliance," in the agency's terminology. Taxpayer contacts through IRP notices and audits potentially improve compliance by informing taxpayers of compliance requirements and by increasing their concern about the penalties associated with noncompliance. Taxpayers who are suspected of noncompliance face the penalty of lost time and other costs associated with the audit itself. Taxpayers whom the IRS determines to be noncompliant may receive civil negligence penalties of 5 percent of the underreported tax or, more rarely, civil fraud penalties of 75 percent plus half the interest due (26 U.S.C. 6653). For reasons that will become clearer in Chapter 2, little is known about the extent to which audits encourage compliance, either by informing taxpayers or by stimulating fear of these penalties. In addition to routine examinations, the IRS has established special programs that use other investigatory techniques to focus on taxpayers in known problem areas, such as construction contractors specializing in home improvements and recipients of tip income. These programs led to the examination of over 26,000 returns and assessed almost $140 million in additional taxes in 1985 (IRS, 1986a). Other recent special programs involving individual tax returns have focused on those who have purchased tax shelters deemed abusive by the IRS (141,000 examinations in 1985 produced $2.5 billion), tax protesters, and those who overclaim exemptions on W-4 forms. Criminal penalties for evasion and other tax offenses, which are sought and imposed very rarely, serve two primary purposes. One, carried out under the General Enforcement Program, is to publicly condemn tax evasion and deter it by punishing the most egregious offenders. The sec-

32

Paying Taxes

ond, carried out under the Special Enforcement Program, is primarily to augment the punishment and investigative capability available with regard to an offender who is charged with another federal crime, by adding tax evasion as a secondary charge. Criminal penalties, unlike the civil penalties imposed by the IRS, require a court conviction and involve other institutions, primarily the U.S. Department of Justice. They also usually require extensive preindictment investigations, which are carried out by the IRS Criminal Investigation Division. These costs keep the annual number of criminal sentences for Internal Revenue Code violations quite small— 1,515 cases excluding narcotics-related issues in 1985. The Criminal Investigation Division sometimes undertakes special projects, investigating groups of persons who are suspected of similar evasive activity (e.g., fraudulent refund claims by prisoners, failure by banks to file reports of large cash transactions, understatement of cash receipts by catfish farmers). On the basis of case studies of several such investigations (IRS, 1978b), the division concluded that the widely publicized arrests, indictments, trials, and convictions that followed these investigations appeared to produce at least temporary decreases in levels of these activities. Recently the Criminal Investigation Division, reflecting general IRS policy, has focused on cases involving illegal tax protesters (302 convictions) and fraudulent tax shelters (55 convictions). To date, the heavy requirements for investigative and prosecutorial resources have constrained criminal prosecutions for use as a narrowly focused deterrence weapon. Considerable research would be needed to learn whether public information about criminal investigations and prosecutions encourages compliance by taxpayers who are unlikely to become targets of such efforts. D E C R E A S I N G THE COSTS OF COMPLIANCE

A number of I R S activities are aimed at decreasing the costs imposed on taxpayers in determining and reporting tax liability. The IRS expects lower cost to decrease the likelihood that taxpayers will simply fail to file or fail to report a particular item because of the complexity involved. The IRS tries to reduce compliance cost by simplifying tax forms and instructions, by providing publications and telephone-answering services, and by offering personal help to assist in filing and to resolve problems taxpayers may encounter in dealing with the IRS. There are no reliable estimates of how these activities affect compliance. The Tax Forms and Publication Division of the IRS aims to develop forms that are simple and likely to lead to accurate calculations. The devel-

Paying Taxes

33

opment of the simplified 1 0 4 0 E Z form is only one of many changes made to simplify forms during the past decade. As part of this program, comments about forms are encouraged through hearings and requests to affected groups and tax practitioners. Procedures for field testing hypothetical problems on sample audiences have been developed recently. The success of simplification efforts is limited by annual deadlines for incorporating the large number of changes imposed by amendments to the tax code: 200 of the division's 375 forms are revised on an annual basis (IRS, 1985^23). In addition to this time pressure, a tax form coordinating committee must approve all changes. Representatives from other divisions on the committee are usually more concerned with other enforcement aspects of the form, such as accuracy of interpretation or the usefulness of the information requested for enforcement. Other efforts to simplify filing requirements include research into a magnetic tape filing system, which would reduce the number of individuals required to file returns on paper. The Taxpayer Services Division provides assistance to the public about the requirements of the tax law and about the status of a taxpayer's account. During 1987, the I R S responded to about 54 million calls, including nearly 11 million handled by a recently installed system of recorded messages (IRS, 1987). An internal probe found an overall accuracy rate of 93 percent on these calls. However, the U.S. General Accounting Office (1987a) reported an accuracy rate of only 79 percent in its probe, and further study of the matter is planned. Following widespread complaints about taxpayer confusion following the Tax Reform Act of 1986, the I R S announced that it would not assess penalties for noncompliance that taxpayers could demonstrate occurred because of errors by taxpayer service representatives. In 1987, the Taxpayer Services Division offered personal assistance in 411 permanent locations, and the I R S voluntary income tax assistance (VITA) and tax counseling for the elderly programs trained about 57,000 volunteers who assisted 1.7 million taxpayers. Finally, 8.6 million questions about taxpayers' accounts were answered in 1987. Although rough estimates of the compliance effects of these activities could probably be obtained using fairly standard research techniques, the panel is aware of only one planned attempt to do so. The I R S and its chief counsel provide written interpretations of the tax code through regulations, revenue rulings, and private letter rulings, which affect most taxpayers primarily by informing tax practitioners. In addition, over 2 million technical referrals a year are handled through taxpayer services. Finally, the problem resolution program intervenes to resolve

34

Paying Taxes

errors and misunderstandings between the IRS and the taxpayer, resolving 407,000 problems for taxpayers in 1987. Although the General Accounting Office (U.S. GAO, 1987b) found that special assistance through the problem resolution program increased levels of taxpayer satisfaction, it also expressed concern about the quality of the feedback the IRS obtains through its efforts to evaluate the program. What is more important from the panel's perspective, very little is known about the effects of these programs on compliance. E N C O U R A G I N G COMPLIANCE T H R O U G H PUBLIC COMMUNICATIONS

The IRS has increased its efforts to encourage compliance through public communications. In 1987 outreach and educational efforts included the distribution of materials to nearly 4 million students, the training of 4,000 teachers, and the distribution of six 15-minute films through the "Understanding Taxes" program. In addition, the IRS and the General Counsel's office organized or participated in workshops for 47,000 small business owners, institutes for 36,000 tax practitioners, and student tax clinics at twelve law and graduate accounting schools. The IRS participated in seventeen televison and seven radio tax clinics during 1987, with estimated total audiences of over 17 million. In total the IRS received about $45 million in free advertising from publications, radio, and television (IRS, 1987). The Public Affairs Division constandy deals with the media and professional organizations, particularly to inform them of changes in laws, regulations, and practice. Little is known about the role of communication in encouraging compliance, and considerable investigation is needed to learn how to combine communication strategies effectively with enforcement programs. REGULATING T A X PRACTITIONERS

Tax practitioners—lawyers, accountants, and others who prepare tax returns or advise clients about tax matters—provide an important component of the tax system. Since they prepare nearly half of all individual tax returns each year, they affect a much larger proportion of returns than IRS examiners. They are more easily monitored than taxpayers since they are a much smaller group. In many ways, they are more vulnerable to punitive actions by the IRS or by professional organizations to which they belong, since the loss of credentials and reputation could eliminate their source of livelihood. Good relations with IRS officials are probably useful to their work. They may probe for unreported income, question implausible

Paying Taxes

35

claims, and encourage compliance by their clients in other ways. Or, by encouraging taxpayers to play the odds against being audited or hinting at what is difficult for auditors to detect, they may encourage noncompliance. Only recently have researchers begun to study how compliance is affected by return preparers and by other practitioners who offer advice and structure transactons so as to maximize after-tax income. One source of IRS control in this area comes from its authority to administer and enforce regulations governing administrative proceedings within the IRS. Although attorneys and certified public accountants are automatically authorized to practice by the Agency Practice Act of 1965, over 31,000 other "enrolled agents" have been authorized by the IRS. Their duties and responsibilities are governed by Treasury Circular 230, and they can be disciplined by suspension or disbarment for such acts as false and misleading advertisement, bribery, submission of fradulent returns, or fraudulent opinions relating to abusive tax shelters. Legislation over the years has provided penalties for preparers, many of whom are not lawyers, certified public accountants, or enrolled agents. These provisions provide civil penalties for abuses such as not signing a return, filing false returns, and so on. Abuses are discovered during routine examinations when preparers are discovered to be involved with an incorrect return. Files kept in district offices and service centers are used to identify consistent abusers, whose other clients' returns are then flagged for examination. In the late 1970s, about 5 percent of all examinations were related to this practitioner program (U.S. GAO, 1976), although a cost-saving decision not to record preparers' identification numbers on computer files in the early 1980s decreased the IRS's ability to request examinations in this way (U.S. GAO, 1982). ACTIVITIES OF THE I R S RESEARCH DIVISION

Although several IRS divisions carry out research and related activities, the center for guiding most research on compliance by individual taxpayers is the IRS Research Division. The Research Division employs a staff of about sixty professionals and controls an annual budget of just over $4 million for purchasing equipment, supplies, and services under the general heading of research and development. Its functions may be classified in five categories. 1. Taxpayer Compliance Measurement Program coordination. The Research Division plans and coordinates the entire TCMP program.5 This coordination, which requires two to three professional staff years annually, is only

36

Paying Taxes

a small fraction of the full cost of TCMP, which the IRS estimates at $128 million for the 1985 survey.6 2. Compliance measurement. Using TCMP and other data sources, the Research Division prepares and updates measures of the amount and composition of the noncompliance tax gap. This effort requires seven to eight staff years plus a small contract budget. 3. Management studies. This category includes workload projection, resource allocation, development and testing of computer-based strategies for detecting unreported taxable income, DIF development, expert systems development, evaluation of forms simplification efforts, and analysis of specific statistical problems encountered in tax administration. These efforts require about forty-two staff years annually, plus perhaps $2 million in contractual costs. 4. Surveys. Surveys of taxpayers and, recently, tax practitioners require about two staff years of Research Division personnel, plus about $500,000 in contractual costs. 5. Liaison with the research community. Although such activities as annual research conferences, responses to researchers' requests for data and questions concerning interpretation, and others consume only one to two staff years of Research Division time, they require programming and legal support from other parts of the IRS and perhaps $500,000 per year in contractual costs. These figures make clear that TCMP dominates the IRS investment in individual taxpayer compliance research, costing about sixteen times as much as all other Research Division activities. The primary justification for TCMP is its use in development of the rules for selecting returns for audit; these rules have been demonstrated to improve the net yield from audits by far more than the TCMP program cost. Similarly, most of the projects in the management studies category are either efforts to ensure that staffing is adequate in all locations or tests of enforcement innovations intended to increase the net cost effectiveness of enforcement activity. By reducing the incidence of breakdowns in processing returns, appeals, and refunds and by discriminating between effective and ineffective programs for revenue recovery, these studies pay for themselves in quite visible ways. The benefits of long-range compliance research activities—compliance measurement, taxpayer surveys, and liaison with the research community—are less visible and more speculative. Therefore, it is not surprising that the investment in these activities is less than 2 percent of the cost of TCMP and the management studies.

Paying Taxes

37

Policy Analysis and Social Science Not all successful social science research on taxpayer compliance will be useful to the IRS policy makers who operate these programs or to others in government who legislate or carry out tax policy. Nor should it be. But there are some natural areas of overlap between the policy concerns of the tax community and the intellectual questions that intrigue scholars. In discussing potential areas of overlap, it is useful to begin by listing some of the considerations that will be weighed by tax administrators considering a change in program (i.e., in tax laws or administrative procedures): 1. IRS I administrative costs. Changes in IRS resources required to set up and implement a new program. 2. Private-sector costs. Changes in the costs incurred by taxpayers, thirdparty information reporters, and tax practitioners for such activities as the maintenance of records, the filing of reports and returns, and the time involved in IRS contacts and appeals. 3. Intangible social costs. Changes in taxpayers' perceptions of the intrusiveness of enforcement—the need to keep records to document legitimately deductible expenses, reluctance to entrust the government with personal financial information, fear and psychic costs evoked by contacts with the IRS, awareness of being watched by an impersonal and mysterious bureaucracy. 4. Political costs and constraints. Reactions by the public, Congress, public interest groups, and budget makers that may thwart a program. Examples include reactions against the 1984 statute and subsequent regulations requiring contemporaneous record keeping for vehicles used for business and pleasure, reactions against laws requiring banks to withhold taxes due on interest paid, recurrent battles to increase IRS computer capabilities and budget. 5. Direct revenue effects. The revenue yielded directly by new or expanded enforcement programs (e.g., audits, matching of information returns, new computer technologies) or the net revenue effects of reallocating resources from one enforcement activity to another. 6. Compliance-based revenue effects. The revenue effects of policy changes that occur indirectly, through changes in compliance. For example, a well-publicized shift in criminal investigations from catfish farmers to promoters of abusive tax shelters may deter investments in such shelters but encourage noncompliance by catfish farmers. A decision to waive penalties incurred because of mistaken advice from a taxpayer services representative

38

Paying Taxes

may strike some taxpayers as fair but aggravate others' resentment of complexity. A tax amnesty may induce some taxpayers to clear their consciences, while angering those who have routinely complied and causing others to question their perception of the IRS as an omniscient tax collector. Of these six considerations, the last is the one for which social science can be most helpful. As the examples of compliance-based revenue effects were selected to show, policy instruments tend to be swords with at least two edges. The diverse theories that underlie the various branches of social science can call attention to these edges, and empirical research can measure their sometimes conflicting effects on compliance. At present much of the potential of social science for informing legislators and policy makers about the compliance effects of their actions remains unrealized. There are several reasons. First, taxpayer compliance research is quite recent (i.e., of the last ten to fifteen years) because, outside the IRS, the subject has received extensive research attention only since the existence of a large purported underground economy in the United States was publicized in the late 1970s. Second, although researchers have begun to cross traditional disciplinary boundaries in the last few years, most existing research is quite fragmented. As a result, a central concern of one discipline—which may reveal the second edge of a policy sword—may be ignored by a researcher from another. Third, while researchers have discovered a variety of statistically significant correlations between compliance and other variables, very little can be concluded about the magnitudes of the compliance differences found. In part, this imprecision exists because the field is new. But it also reflects the difficulty of measuring compliance and of gaining access to recent useful data—some of the most recent research is based on data for 1969. And it reflects the fact that the IRS frequently fails to take advantage of changes—in the law, in administrative procedures, and in resource allotments—as opportunities to measure compliance responses. Fourth, research has only begun to progress in reflecting important institutional realities of taxpaying. This beginning reflects progress on two fronts: the development of tractable theoretical and empirical models that can cope with such complexities and enhanced communication among the communities of tax administrators, tax practitioners, and social scientists. In developing its recommendations, the panel has tried to provide an agenda that balances the concerns of immediate policy relevance and those of long-term understanding, that respects the privacy concerns and rights

Paying Taxes

39

of taxpayers, and that promises results commensurate with the administrative costs and political risks involved. In suggesting research we have tried to ensure that it is within acceptable bounds for equal treatment of classes of taxpayers and that it is likely to inform policy debates. Our research agenda and recommendations for its execution are influenced by additional organizational priorities and constraints extending beyond data-base development and disclosure. Of fundamental importance are the differences between organizations such as the National Science Foundation, established to support basic or general applied research, and the research arms of mission agencies such as the IRS. The former organizations can be guided primarily by assessments of likely scientific payoffs from proposed research. However, the latter must also weigh the prospects of visible contributions and costs in terms of the agency's primary mission. This basic difference will sometimes lead the two kinds of organizations to place different priorities on a given research question. It is therefore incumbent on advocates of particular lines of research by mission agencies to make their case primarily in terms of their practical policy implications. Even when a research question is acknowledged to have broader scientific implications, the research design will be influenced by more immediate considerations. For example, in assessing the feasibility of a particular randomized experiment, the I R S must consider whether it will stimulate legal challenges, ethical objections, overt public opposition to the agency (with possible adverse effects on compliance), or less visible social delegitimization of the agency that could aggravate noncompliance. For these reasons, in developing its recommendations the panel attempted to emphasize questions having both scientific and policy interest, to explain the practical significance of the proposed research, and to advocate approaches yielding generalizable results while recognizing organizational and resource constraints.

Plan of the Report In this chapter, we have offered a definition of compliance with federal income tax reporting requirements and discussed available estimates of the magnitude of noncompliance. In Chapter 2, we review and synthesize previous theoretical and empirical research on factors that may affect taxpayer compliance, and we recommend promising extensions of several existing lines of work. In Chapter 3, we argue that several new lines of

40

Paying Taxes

research should be opened, to study how compliance is affected by the social and financial environments that people find themselves in, and by the mental processes through which they make sense of these complex and uncertain environments. These matters suggest that various components of the tax system itself may be important influences on compliance and that they should receive more research attention than they have to date. To illustrate the point, in Chapter 4 we suggest useful research on how three components—legislative and administrative change, tax practitioners, and IRS taxpayer contact programs—affect compliance. In Chapter 5 we recommend a number of innovations for improving the quality of compliance research data and making it more accessible to compliance researchers, while protecting the rights of citizens as taxpayers and research subjects and the security of confidential IRS enforcement information. Finally, in Chapter 6 we argue that knowledge about taxpayer compliance can best be accumulated by a diverse group of independent researchers who communicate actively among themselves and with the IRS, and we recommend a number of institutional innovations to facilitate interaction and cooperation.

Supplementary Notes A : PATTERNS AND T R E N D S IN TAXPAYER COMPLIANCE

The panel's interest in compliance measurement was primarily in the compliance of individuals. Measures of aggregate noncompliance were of some interest, however, as a context for our work. This note focuses primarily on the aggregate noncompliance estimates that were available when the panel began its work at the end of 1984. While these studies were concerned with different time periods, many of them happened to include estimates for the year 1976; therefore, for comparability, much of the following discussion focuses on that year. At the outset it should be emphasized that the estimates reported in this note are subject to very large potential errors. In addition, seemingly trivial differences in concept definitions or in the wording of survey questions can increase or decrease estimates by factors of 2 or 3. Chapter 5 of this report contains several recommendations for improving the individual-level compliance measures to be used in future empirical research. Some of those recommendations, if implemented, might eventually improve the accuracy of the aggregate estimates. But the panel did not critically review the aggregate estimates nor attempt to develop recommendations to improve them.

Paying Taxes

41

In interpreting aggregate noncompliance estimates, at least five different concepts should be distinguished. The first two are concerned with taxable income, and the rest are concerned with income tax. The first concept is the so-called underground economy, the value of all economic activity that is unrecorded in national income accounts and is presumably untaxed. Estimates of the size of the underground economy were made by Feige (1979, 1980), Gutmann (1977), Henry (1975, 1976, 1983), and Tanzi (1980, 1982). Although their studies differ in details and underlying assumptions, the basic strategy is to count money in circulation, estimate the share of that money used in recorded transactions, and to assume that the rest is being used in unrecorded activity. As shown in Table 1, the 1976 estimates differ substantially, from Henry's range of $40-865 billion to Feige's estimate of $369 billion annually. All these estimates are analogous to the national income accounting concept of gross national product, because they are intended to include the value of all goods and services produced but not recorded. But the outputs of some producers in the underground economy are sold to other producers in intermediate transactions; if these other producers were reporting their income for tax purposes, they would first deduct intermediate transactions as allowable business expenses. Consequently the underground economy concept is overly inclusive as an indicator of income that is untaxed because of noncompliance with federal income tax laws. The second concept, unreported taxable personal income, excludes the value of intermediate transactions and income earned by organizations rather than individuals. As shown in Table 1, two estimation approaches have been used. Kurtz and Pechman (1982) have derived an indirect estimate for 1976 of $78 billion from discrepancies in personal income accounting statistics. The IRS (1979b, i983f) has compiled direct estimates of between $100 and $154 billion by adding up estimates of components (e.g., unreported tips, overstated deductions, failure to file returns) obtained using a variety of methods. The third concept, the income tax noncompliance gap, is a measure of the tax revenue lost because of all forms of noncompliance by taxpayers. Estimates by the IRS (i983f) for several years, summarized in Table 2, amount to about $43 billion for 1976 and about $90 billion for 1981. The gap includes the taxes due on unreported taxable personal income plus two other components: the corporation tax gap and the remittance gap—the failure of employers to deposit all taxes withheld on their employees and the failure of taxpayers to remit all taxes due along with their returns. The fourth concept, the individual income tax reporting gap, conforms

42

Paying Taxes

TABLE 1

Estimates of Aggregate Unreported Taxable Personal Income for

1976

Concept and Source

Originally published ($ billions)

AdjustedJ ($ billions)

Underground economy Feige (1979, 1980) Gutmann (1977) Tanzi (1980, 1982) Henry (1975, 1976, 1978)

369 176-240 138-199 40-65

157 61-93 78-124 72

12 5-7 6-9 5

77

6

61-96 119

5-7 9

Unreported taxable personal income Indirect estimates: Kurtz and Pechman (1982) 78 Direct estimates: IRS (1979b) 100-135 IRS (1983f) 154

Percent of personal incorni

Reflects adjustments by Henry ( 1 9 8 3 ) t o achieve conceptual uniformity and further adjustment for the panel's use to remove illegal income, estimated at $ 2 6 million by Simon and W i t t e ( 1 9 7 9 ) . b 1 9 7 6 personal income estimate o f $ 1 , 3 5 1 billion is f r o m Bureau o f Economic Analysis ( 1 9 8 1 , Table 8 . 7 ) . Source: In all estimates except I R S ( 1 9 8 3 f ) , originally published estimates are taken f r o m Henry ( 1 9 8 3 ) . I R S ( 1 9 8 3 f ) originally published estimates are the sum o f filers' unreported income (Table III-1, p. 1 2 ) and nonfilers' balance due (Table C - 5 , p. 7 9 ) . a

most closely to the scope of the panel's inquiry. It is computed by subtracting from the estimated income tax noncompliance gap the components associated with corporations, nonremittance, and illegal activity. As shown in Table 2, the I R S (i983f) estimates that the reporting gap amounts to about 75 pecent of the overall income tax noncompliance gap—$32 billion in 1976 and $68 billion in 1981. The fifth concept, the net individual income tax gap, is the portion of the individual income tax reporting gap that remains after I R S administrative actions. It was estimated by the I R S (1986b) for 1981 by subtracting from the individual gap the estimated $8—$xi billion of revenue yielded directiy by checking arithmetic errors, matching tax returns to information reports, auditing returns, collection activities, and criminal investigations. This estimate, of course, takes no account of secondary or "ripple" effects of these activities on compliance by taxpayers who are not contacted directly through these programs. Compared with a 1981 gross (preenforcement)

Paying Taxes TABLE 2

43

I n c o m e T a x G a p , 1 9 7 3 - 1 9 8 1 ( $ billions)

Legal-sector tax gap Total Corporation tax gap, total Individual tax gap, total Individual income tax reporting gap: Total Nonfilers' income tax liability (Net o f prepayments and credits) Filers' income tax liability: Unreported income Overstated business expenses Overstated personal deductions 3 Net math error Individual income tax remittance gap, total: Employer underdeposit o f withholding b Individual balance due after remittance Illegal-sector tax gap (partial)'

1973

1976

1979

1981

28.8 3.5 25.3

39.2 4.6 34.6

62.3 6.4 55.9

81.5 6.2 75.3

23.8

32.2

50.6

68.5

0.9 22.9 17.3 2.1 3.4 0.1

1.4 30.8 24.2 3.4 3.0 0.2

2.0 48.6 38.4 4.7 5.0 0.5

2.9 65.6 52.2 6.3 6.6 0.5

1.5

2.4

5.3

6.8

1.1

0.9

1.8

2.4

0.4 2.1 (0.8)

1.5 3.4 (1.3)

3.5 6.3 (2.2)

4.4 9.0 (3.2)

Includes itemized deductions, personal exemptions, and statutory adjustments. Includes a small amount for underreported withholding by employees and a small negative amount for undcrclaimed withholding bv individuals. c Includes income from illegal drugs, illegal gambling, and prostitution only. Figures in parentheses are standard errors. Source: I R S ( 1 9 8 3 f , T a b l e 1-1, p. 3 ) .

h

individual income tax g a p o f $75 billion, the net gap was estimated at $ 6 4 — $ 6 7 billion. THE UNDERGROUND ECONOMY AND UNREPORTED TAXABLE PERSONAL INCOME.

In the absence o f any means o f direct measurement, estimates o f

underreported taxable income have been produced for a variety o f purposes under a variety o f assumptions. T h e various approaches are described in some detail in H e n r y (1983). Because all o f them are difficult and time consuming, they are not available on a regular basis. M a n y o f the authors report estimates for tax year 1976, and H e n r y (1983) applied the methods o f others to derive estimates for that year. These 1976 estimates are summarized in Table 1. T h e table reports both the authors' original estimates and

44

Paying Taxes

estimates adjusted by Henry that are intended to compensate for conceptual differences in the figures estimated by the original authors. The estimates summarized in Table i are categorized as "indirect" or "direct," depending on the estimation approach used. All the estimates of the underground economy are indirect, obtained by inferring the immeasurable unreported income from data on measurable quantities— components of the money supply (Feige, 1979, 1980; Gutmann, 1977; Tanzi, 1980, 1982; Henry, 1975, 1976, 1983) or of the national income and product accounts (NIPA) (see Kurtz and Pechman, 1982). Indirect estimates. Indirect estimates of unreported economic activity are derived from discrepancies that emerge in reconciling economic statistics that should be related by definition in the absence of noncompliance. They have been calculated by analyzing monetary data and by comparing adjusted gross income (AGI) totals from individual tax returns with AGI estimates computed by the Bureau of Economic Analysis (BEA) for the national income and product accounts. The latter are based primarily on business tax returns, which are considered more reliable (Parker, 1984). The monetary indirect estimates are based on the presumption that most unreported economic activity takes place in cash. Econometric techniques are applied to data from an earlier period to estimate the relationship between the value of currency in circulation and some other statistic, such as demand deposits (i.e., checking accounts) or national income. The estimated relationship is used to project the expected value of currency for a later period, and the excess of actual currency over expected currency is assumed to be producing unreported taxable income. As noted by Henry (1983), these estimates have been widely criticized on technical grounds—failure to adjust for increased overseas holdings of American dollars, failure to model the effects of increased use of credit cards (which might affect the ratio of currency to checking accounts regardless of trends in unrecorded activity), and failure to test the sensitivity of the estimates to alternative estimates of the velocity of money (i.e., the value of transactions per year per dollar in circulation) and of currency lifetimes, which are themselves subject to great uncertainty. An additional problem in terms of the panel's scope is that the indirect monetary estimates include income earned in illegal activities such as drug trafficking and prostitution; adjustments to remove the illegal components are difficult to make. Finally, the estimates include the income of persons who are not required to file, which should not be included in estimates of noncompliance. The indirect noncompliance estimates based on discrepancies between

Paying Taxes

45

national income account estimates produced by the BEA and personal incomes reported to the IRS have been used as noncompliance benchmarks for many years. They are only approximations, however, because the discrepancies arise from several causes other than taxpayer noncompliance. First, the BEA estimates do not reflect allowable adjustments to income such as moving expenses but do include income earned by persons who are not required to file. Second, BEA approximations to components of income that are not routinely captured by the bureau's data collection systems (e.g., interest to individuals, individual capital gains) are themselves open to question. Direct estimates. The IRS (i983f) direct estimate of unreported taxable income reported in Table 1 includes three components: overstated subtractions from income by return filers ($20.8 billion), unreported income of return filers ($131.5 billion), and net taxable income of nonfilers ($2.1 billion). The estimates of overstated subtractions and some kinds of unreported income for filers are derived from the 1976 cycle of the IRS Taxpayer Compliance Measurement Program, which is described more fully in Supplementary Note B to this chapter. The estimates of nonfilers' taxable income were computed from discrepancies between individuals' IRS master file records and their Social Security Administration and Census Bureau income reports. The IRS considers the TCMP audit a relatively accurate device for measuring overstated subtractions; it is usually straightforward to determine whether the taxpayer has the documentation needed to support an adjustment to income or an itemized deduction. For this reason, unadjusted 1976 TCMP results were used to estimate overstated subtraction items by taxpayers who filed returns. Unreported income is considered more difficult for auditors to establish, because they must first discover its existence. Today, information reports such as forms W-2 and 1099 help to identify unreported wages, interest, dividends, and other categories of income. But in the 1976 TCMP cycle, auditors did not have access to the information reports, and a special study (IRS, 1983^ Appendix B) found that for the income categories subject to information reporting, those reports established far more unreported income than did the TCMP auditors. Although the multiple (the ratio of estimates from information reports to estimates from audits) varied substantially by type of income, unreported taxable income was estimated by multiplying the TCMP estimates for most categories of unreported income (i.e., all categories except tips, informal suppliers' income, and certain types of capital goods, divi-

46

Paying Taxes

dends, alimony, and business income) by a common factor, 3.5. The validity of this adjustment is subject to question, and the most recent estimate (IRS, 1988) is based on a factor of 3.28. Since unreported income accounts for nearly 80 percent of the total estimate of unreported taxable personal income, that estimate should be interpreted with great caution. For some categories of income, TCMP audit results, even adjusted on the basis of information reporting, were not considered sufficiently accurate. Therefore, TCMP-based estimates of unreported income were supplemented by estimates from special studies of certain types of unreported income, including unreported tip income and unreported income of undocumented aliens and informal suppliers, such as street vendors and firewood purveyors with no fixed business address. These special studies are described elsewhere (IRS, i983f). The last major component of the IRS direct estimates for 1976 was the taxable unreported income of citizens who failed to file tax returns as required and therefore could not have been selected for TCMP. Nonfiler estimates were obtained using "exact match" data files for 1972 and 1977 that contained information about individuals from the IRS master file (including an indicator whether a return had been filed), the Census Bureau's Current Population Survey, and Social Security Administration administrative records. Individuals with Social Security or Census Bureau records indicating receipt of taxable income, but no matching IRS record, were flagged as potential nonfilers. Their income levels and other relevant information were checked to see if they should have filed a tax return. For individuals who apparently should have filed on the basis of their income and employment status, estimates of taxable income were added to the underreporting estimates for nonfilers. Over 4.2 million nonfilers were estimated for 1976. However, nearly 1.8 million of them would have received a refund or owed no additional tax because of prior withholding. Therefore, only about 2.4 million nonfilers avoided tax liability, and the net contribution of nonfiling to total unexpected income was only $1.4 billion. Questions have been raised concerning errors in the construction of the exact match files, and these errors may have created inaccuracies in the estimates of unreported income by nonfilers. Another commonly reported measure is the noncompliance tax gap—the tax loss associated with unreported taxable income and uncollected tax by individuals and corporations. This measure has been calculated for tax year 1976 (IRS, i983f) and projected forward to 1981 under the assumption that taxpayers reported constant fractions of the components of taxable income and subtracT H E I N C O M E TAX N O N C O M P L I A N C E

GAP.

Paying Taxes

47

tions throughout that period. Both the 1976 and 1981 tax gap estimates are reported in Table 2, which is reproduced from IRS (i983f:3). The 1981 estimates are subject not only to the problems discussed above in estimating the components of unreported income for 1976, but also to errors that may stem from changes in compliance behavior between 1976 and 1981. According to the IRS estimates for 1981, about $90.5 billion, or 20 percent, of corporate and individual income tax liability was not collected without IRS intervention. Of this amount, corporate returns accounted for $6.2 billion, and returns involving illegal income accounted for $9 billion. Both measures involve major problems that undermine their validity but are not considered further because corporate and illegal sector compliance are beyond the scope of this report. Another $6.8 billion, the remittance gap, was reported but not paid. T H E I N D I V I D U A L I N C O M E TAX R E P O R T I N G GAP. The remaining amount in Table 2, a 1976 individual income tax reporting gap of $32.2 billion, measures the concept of noncompliance that corresponds most closely with the scope of the panel's inquiry (IRS, i983f). It represents about 75 percent of the entire estimated income tax gap. For 1981 the estimated reporting gap was $68.5 billion, of which $65.6 billion was attributed to inaccurate filed returns and the rest to nonfilers. This amounts to an estimated 17 percent of true tax liability. While it is important not to attribute great precision to the estimates, they suggest that somewhat more than 80 percent of true tax liability is accurately reported by compliant taxpayers without being contacted by the IRS. Nevertheless, the 1981 noncompliance tax gap amounts to approximately 87 percent of the federal deficit in that year and 37 percent of the average annual deficit since then. As this report was in final preparation, the IRS (1988) released new estimates of the individual income tax reporting gap over the period 19731992. The latest estimates are said to be more accurate than those in IRS (i983f) because of refinements in the adjustments for unreported income not discovered by TCMP auditors and in the computations of the tax rates that apply to the unreported income. Unfortunately, IRS (1988) does not contain any estimate for filers only for 1981, but the estimate for 1982 is only $51.9 billion, of which $46.2 billion was attributed to filed returns— lower than the 1981 estimates in IRS (1983^. IRS (1988) projects an individual income tax reporting gap of $79.3 billion for 1986, a decrease to $63.5 billion in 1987 because of the Tax Reform Act of 1986, and a trend upward to $82.6 billion in 1991, not adjusted for inflation. THE NET TAX GAP. TO complete the story, we need to consider how

48

Paying Taxes

much of this noncompliance is caught and corrected through IRS enforcement actions. A recent study of the net tax gap (IRS, 1986b) estimated that the program for securing returns from delinquent nonfilers caught $287 million, or 10 percent, of the $2.9 billion in unreported taxes from nonfilers in 1981 and yielded an additional $66 million in interest and penalties. Of the $65.6 billion tax-reporting gap for individuals who filed in 1981, $458 million in taxes due (0.7 percent) was discovered by the computerized arithemetic error checking program, $490 million (0.7 percent) by matching against information reports, and $2.7 billion (4.1 percent) in the examination program. Thus, only an estimated 6 percent of the $68.5 billion tax gap from individual reporting errors was discovered through various enforcement programs. The same study notes that "about $28 billion could potentially have been recovered with current sources of information and current IRS enforcement programs funded for operation at their highest economically feasible levels" (IRS, i986b:i; the figure includes returns from collections activity and amounts to about 31 percent of the $10.5 billion noncompliance). Although the assumptions required to generate such an estimate preclude putting much confidence in it, this estimate suggests that at best only something like 30 percent of the total tax gap could be collected efficiendy by enforcement actions, and much of that would be collected from corporations. The relatively small percentage of the individual tax gap that is actually caught by enforcement actions underscores the potential importance of compliance, in contrast to the revenue yield from enforcement, in reducing the tax gap. OTHER N O N C O M P L I A N C E M E A S U R E S . Although estimates of aggregate noncompliance are useful for describing the magnitude of the tax compliance problem, other statistics can provide additional insights into the nature of compliance and can further inform the development of policies for encouraging compliance. The IRS regularly reports additional data on patterns of noncompliance: the prevalence of noncompliers in the population, the extent of certain disaggregated categories of noncompliance for specific types of taxpayers, and trends in noncompliance of certain types. These data are described in the following sections. Prevalence of noncompliers. After presenting estimates of the noncompliance tax gap, it is logical to ask whether the gap is accounted for by a few taxpayers, each of whom underreports a large amount of taxable income, or whether a broad cross-section of the taxpaying population underreports by small amounts. The fraction of the population that fails to comply may be called the prevalence of noncompliers. Estimates of noncomplier preva-

Paying Taxes

49

lence are available from a variety of sources. As a first step in reporting these estimates, it is important to distinguish between current prevalence—the fraction that fails to comply in a single tax year—and cumulative prevalence—the fraction that fails to comply at least once during an observation period of several tax years. Because a multiyear observation period presents more opportunities to fail to comply, cumulative prevalence should always equal or exceed current prevalence for any group of taxpayers. Moreover, cumulative prevalence should increase with longer observation periods. However, these distinctions have usually been ignored in comparisons of prevalence statistics across studies (see, e.g., Westat, Inc., i98of; Kinsey, 1984).

Estimates of both current and cumulative noncomplier prevalence can theoretically be based either on the results of I R S audits or on self-reports of surveyed taxpayers who are asked whether they have failed to comply. However, even leaving aside measurement errors arising from such sources as auditors' errors and survey respondents' distortions, prevalence estimates derived in these two ways are not stricdy comparable. Even candid survey respondents can acknowledge only noncompliance of which they are aware, and so will fail to report noncompliance arising from complexity or misunderstandings, unless the instances are brought to their attention by the I R S . In contrast, prevalence statistics based on T C M P audits reflect noncompliance for any reason and should therefore be larger than surveybased estimates in any sample, even if survey respondents have no memory lapses and make no effect to conceal noncompliance. In addition to this conceptual difference, operational problems have so far made it impossible to produce comparable prevalence estimates from T C M P and taxpayer surveys. With the exception of a panel study involving small samples of taxpayers between 1969 and 1973, T C M P cycles have involved cross-sectional samples whose tax returns for only a single year are examined. Because a sample of taxpayers cannot be repeatedy audited without affecting compliance behavior, the T C M P program as currently structured cannot produce estimates of cumulative prevalence over a multiyear observation period. In contrast, survey researchers, probably as a means of reducing respondent sensitivity to questions about noncompliance, have nearly always probed for noncompliance during such multiyear periods. Consequently, except for the 1978 current prevalence estimates of Aitken and Bonneville (1980), all surveys of noncompliance have produced only cumulative prevalence estimates, which are of course not comparable.

50

Paying Taxes Audit-based

noncomplier

prevalence

estimates for

1979.

T a b l e 3 presents

1 9 7 9 prevalence estimates f o r nonfilers, misreporters o f i n c o m e , a n d misreporters o f subtractions f r o m taxable i n c o m e . T h e estimates f o r nonfilers w e r e c o m p u t e d in the exact m a t c h s t u d y previously discussed, a n d the estimates f o r filers are d e r i v e d f r o m the 1 9 7 9 T C M P survey. T a b l e 3 indicates that a c c o r d i n g t o I R S

( 1 9 8 3 ^ , 5.6 percent o f all

p e r s o n s w h o w e r e r e q u i r e d t o file a 1 9 7 9 tax return failed t o d o so. T h e m a j o r i t y o f these, 3.2 percent, if t h e y had filed, w o u l d h a v e s h o w n a balance

TABLE 3 Prevalence o f Noncompliant Taxpayers based on T C M P Measures, 1 9 7 9 (percentage) Current

Type of noncompliance

All filing taxpayers

prevalence Taxpayers meeting preconditions'

Failure to file return0 W i t h balance due With refund o r zero balance Misreporting

NA NA

3.2 2.4

incomec

Underreporting taxable income

42.3

42.3

6.0

6.0

Underreporting taxable income

4.4

32.2

Overreporting taxable income

2.5

18.5

17.5

61.0

7.8

27.1

Overreporting taxable income Misreporting

subtractions from incomec

Adjustments to income resulting in c

Itemized deductions resulting in c Underreporting taxable income Overreporting taxable income

NA = not applicable. ' Taxpayer base is restricted to exclude taxpayers who could not have engaged in a specific form o f noncompliance. For rows 1,2,3a and 3b, the respective bases are: all persons required to file, all filers reporting income, all filers reporting adjustments to income, and all filers itemizing deductions. b Source: Nonfiler estimates from IRS (1983f, Table C-4, p. 78) include delinquent nonfilers plus enforcement-secured returns. Base includes nonfiler estimates plus count o f files. c Source: American Bar Foundation analyses of 1979 T C M P data files prepared by the I R S Research Division. Item misclassifications that do not affect tax liability are excluded from counts o f noncompliers. Base for "all taxpayers" measure is filed returns. Adjustments include subtractions such as moving and business expenses and payments to IRA or Keogh retirement accounts. Itemized deductions include allowable medical, interest, charitable, and other Schedule A expenses.

Paying Taxes

51

due despite withholding and prepayments during the tax year. But a sizable minority, 2.4 percent, would have shown a refund due. Table 3 also presents 1979 prevalence estimates for specific forms of noncompliance on filed returns. These estimates are abstracted for the panel from K. W. Smith's (1985) tabulations of 1979 TCMP data. According to those data over 50 percent of all returns correctly reported all income items; of the 48 percent with reporting errors, 42 percent involved underreporting taxable income and 6 percent involved overreporting. Prevalence estimates are also reported in Table 3 for taxpayers who overand underreport two kinds of subtractions from income: adjustments to income, such as employee business expenses, and itemized deductions. In interpreting these statistics, it is important to note that the majority of returns filed do not involve these items. In 1979 only 14 percent of returns included adjustments and 29 percent itemized deductions. Therefore, the fraction of noncompliers is reported both as a fraction of all filing taxpayers and as a fraction of taxpayers meeting preconditions (e.g., the number who failed to adjust income accurately is divided by the number for whom adjustments were an issue, according to either the taxpayer or the TCMP auditor).7 It is significant that of the taxpayers meeting preconditions, 18 percent misreported adjustments to income and 27 percent misreported itemized deductions in such a way as to overstate their taxable income. These large figures suggest that, relative to reporting income, subtracting allowable adjustments is even more subject to errors and misunderstandings. Noncomplier prevalence by return item. Table 4, which is also based on tabulations of 1979 TCMP data by Smith (1985), reports noncomplier prevalence by return item, for all items that were used, or that auditors thought should be used, on at least 2 percent of all filed returns. Each prevalence estimate is expressed as a fraction of taxpayers who met preconditions, i.e., taxpayers who either reported a nonzero amount for the return item or should have done so according to the TCMP auditor. (The fractions meeting preconditions, which also provide some indication of the policy significance of noncompliance on the specific return items, are reported in the last column of the table.) Return items are grouped into categories of income and subtractions from income and are listed in order of the prevalence of taxpayers who underreport taxable income. Three sets of prevalence estimates are reported: the fraction that underreported taxable income by any amount, the fraction that underreported taxable income by more than $50, and the fraction that overreported

52

Paying Taxes

TABLE 4

Noncomplier Prevalence by Return Item (percentages) Underreporting taxable income?

Return

Any amount

$51 +

Overreporting taxable income

Meeting preconditions

Income items Total Farm income (Sch. F) Proprietor's income (Sch. C) Rents/royalties Tips Interest Other income (Sch.

42.3 66.9 65.2

NA 62.5 60.3

6.0 14.3 11.1

100.0 2.8 10.9

47.4 45.4 39.8 29.9

42.1 29.9 14.2 23.7

14.1 1.4 3.3 4.1

9.0 3.1 79.3 10.0

E) Capital gains (Sch. D) Partnership income Taxable dividends State/local tax refund Wages/salaries

26.7 26.4 25.8 18.3 6.6

22.1 25.0 16.5 13.8 5.3

12.1 5.5 8.9 3.3 1.0

10.3 3.2 11.8 15.8 89.2

32.2 61.0

NA NA

18.5 27.1

14.1 28.7

60.5

44.6

24.4

9.4

53.3 46.5

48.7 11.1

7.5 5.5

2.2 3.0

43.9

39.8

22.1

7.6

36.0 33.4 30.6 29.9

27.1 26.1 19.0 13.5

12.7 12.7 20.7 11.1

23.3 25.3 4.4 4.9

29.9

16.3

24.3

28.2

29.8 24.4 18.9 18.5

20.8 17.3 13.7 11.9

19.3 13.7 9.2 7.8

24.3 7.2 25.1 16.8

17.4

11.6

39.2

7.1

Subtraction items Total adjustments Total itemized deductions Nonpremium medical expenses Casualty losses Political contributions credit Employee business expenses Misc. deductions Cash contributions Child care credit Residential energy credit Other state and local taxes Nonmortgage interest Noncash contributions Real estate taxes Medical insurance premiums Investment credit

Paying Taxes Mortgage interest State and local income taxes IRA contributions Self-employment tax Interest penalty (early withdrawal)

11.1 10.1

8.9 7.1

7.5 11.1

22.7 24.1

9.6 8.2 4.2

8.4 NA 2.6

2.8 59.5 29.8

2.6 2.2 2.1

53

Base is returns reporting item per taxpayer's or auditor's interpretation. Numerator includes those found to have illegally reported zero income in category or claimed unentitled subtraction. Source: Derived from American Bar Foundation analysis of 1979 TCMP tabulations. 1

taxable income.8 The pattern is consistent with at least two hypotheses about taxpayer compliance: it is discouraged by greater complexity of calculations and judgments needed to compute the item, and it is encouraged by greater visibility of transactions. For three income items for which noncompliers are very prevalent—farm income, properietor's income, and rents/royalties—taxpayers are expected to understand complicated regulations, to keep detailed records, and to perform fairly extensive calculations. For all three items, some gross receipts are offset by allowable expenses, and noncompliance can occur either because records or the interpretation of regulations fails to support the expenses claimed or because evidence is found of unreported income. In addition, some income in each of those categories may come in the form of cash, which lacks documentation that is immediately visible to the IRS. In contrast, the three income items with the lowest prevalences of noncompliers—dividends, state and local tax refunds, and wages and salaries—are largely precalculated for the taxpayer, and two of the three are made visible to the IRS through information reports (forms W-2 and 1099) from their sources. Compared with income items, subtraction items are more visible to the IRS because the taxpayer carries the burden of documenting their existence. Nevertheless, it may be significant that the five items for which noncompliance is the most widespread involve judgments about the allowability of expenditures, while the five with the least widespread noncompliance leave little to interpretation. Table 4 also demonstrates that overreporting of taxable income occurs for all items and is fairly widespread for some. Since presumably few taxpayers want to pay more taxes than necessary, they may be having difficulty in interpreting complex regulations and performing calculations,

54

Paying Taxes

or they may fail to claim a subtraction that they believe is legal but likely to arouse IRS suspicion. The three income items for which underreporting is most prevalent are also among the items for which overreporting is most prevalent—presumably an effect of the complexities mentioned above with respect to those items.9 The fact that the prevalence of overreporting is generally higher for subtraction items than for income items is consistent with both possible explanations. The regulations governing investment credits, employee business expenses, and the like are complex, and taxpayers may be reluctant to claim such rare deductions because of an expectation that the claim would trigger an IRS audit. Survey-based noncomplier prevalence estimates. Another approach to estimating the prevalence of noncompliance is through surveys that ask respondents if they have failed to comply in the past. Even if survey-based measures were error free, they could be expected to differ from audit-based measures because of auditors' errors that may over- or understate noncompliance, and because of unsettled differences between auditors and taxpayers over complex or ambiguous matters. But surveys are also subject to errors because of noncompliance that respondents do not recognize, forget about, or wish to conceal from interviewers. Very little is known about the magnitudes of these discrepancies, for at least two reasons. First, most surveys measure noncompliance over a reference period of several years and so are not comparable to audit-based measures, which relate to a single year. Second, in the United States, IRS concern over taxpayer privacy has prevented researchers to date from comparing audit and survey results for a common sample of taxpayers. When such a comparison was carried out in the Netherlands, however, Hessing, ElfFers, and Weigel (1986) found essentially a zero correlation. Table 5 summarizes the major survey-based U.S. prevalence estimates. Because of the uncertainties just discussed, the estimates should be treated as very tentative. Most of them are also reported by Kinsey (1984), who provides specific methodological critiques of the estimates. Table 5 summarizes survey-based prevalence estimates for "general" noncompliance—including both understanding income and overclaiming subtractions—and for each of those categories separately.10 Within each section, estimates are listed for the most part in order of reference period for the prevalence estimate from longest to shortest—lifetime, five years, and current (one year). If all the samples represented the same population, the shorter reference periods should of course be associated with smaller estimates. For general noncompliance (Table 5, Section A), survey-based

Paying Taxes Table 5

55

Noncomplier Prevalence Estimates Based «11 Surveys

Prevalence

Study

Prevalence estimate (%)

A. Estimates for general noncompliance Cumulative, Grasmick and Scott lifetime (1982)

25 (1979) 28 (1981)

Cumulative, 5 year

Scott and Grasmick (1981)

20

Cumulative, 5 year

Tittle (1980)

12

Cumulative, 3 year

Minor (1978)

12

B. Estimates for underreporting income Cumulative, Spicer(1974) lifetime

22

Cumulative, lifetime

Yankelovich et al. (1984)

16

Method

Sealed envelope responses of 3 5 0 - 4 0 0 persons interviewed in person in 1979 or 1981 (Oklahoma City) Sealed envelope responses of 329 persons interviewed in person in 1980 (Oklahoma City) Responses to direct questions by 1,993 persons interviewed in person in 1972 (Oregon, Iowa, New Jersey) Responses to direct questions by 274 persons interviewed in person in 1975 (Tallahassee, FL)

Responses to direct questions by 130 household heads interviewed in person in 1974 (Ohio suburbs) Responses to direct questions by 2,208 primary taxpayers interviewed in person in 1983 (United States)

56

Paying Taxes

TABLE 5

(cont.)

Cumulative, lifetime

Westat, Inc. (1980f)

12-15

Cumulative, 5 year

Mason and Lowry (1981)

17

Cumulative, 5 year

Mason, Calvin, and Faulkenberry (1975)

15

Cumulative, 5 year

Habib ( 1 9 8 0 )

12 (randomized response)

Current

Aitken and Bonneville (1980)

2 6 (locked box: how often?)

Responses to direct questions by 3 4 8 primary taxpayers interviewed in person in 1 9 7 9 (San Jose, CA, and Smith Bend, IN). Estimate higher when interview scheduled by appointment. Responses to direct questions by 8 0 1 primary taxpayers interviewed in person in 1 9 8 0 (Oregon) Responses to direct questions by 8 0 0 primary taxpayers interviewed in person in 1 9 7 5 (Oregon) Responses by 172 primary taxpayers interviewed in person by 1 9 8 0 (Multinomah County, O R ) Responses by 3 , 5 8 8 primary taxpayers interviewed in person in 1 9 7 9

21 (randomized response) 12 (locked box: ycs/no) C. Estimates for overstating deductions Cumulative, Spicer(1974) lifetime

25

See entry in Section B .

Paying Taxes Cumulative, lifetime

Habib (1980)

Cumulative, lifetime Cumulative, lifetime Cumulative, 5 year

Yankelovich et al. (1984) Westat, Inc. (1980f) Mason and Lowry (1981) Mason, Calvin, and Faulkenberry (1975) Aitken and Bonneville (1980)

Cumulative, 5 year

Current

16 (randomized response 5 (direct question) 7 6-7 6 5

11 (randomized response) 4 (locked box: yes/no)

57

Sec cntrv in Section B.

See entry B. See entry B. See entry B. See entry B.

in Section in Section in Section in Section

See entry in Section B.

cumulative prevalence estimates follow this pattern: they lie in the ranges of 22—25 percent for a lifetime reference period and 12—20 percent for a five-year reference period. The single study to use a three-year reference period reports a prevalence of 12 percent. For underreporting income (Table 5, Section B), the range of lifetime estimates extends from 12 to 22 percent, and the range of five-year estimates is only slightly smaller, from 12 to 19 percent. Interestingly, the 26 percent single-year prevalence estimate obtained by Aitken and Bonneville (1980) using the locked-box technique is higher than any of the multiyear estimates, most of which were obtained using direct questions. This suggests that sensitive question techniques have at least partial success in eliciting admissions of noncompliance. However, even this estimate is well below the 42 percent estimate derived from the 1979 TCMP and reported in Table 4. For overstating deductions, all but one of the cumulative estimates obtained using direct questions lie in a lower range, 5 to 7 percent of all respondents. The Aitken-Bonneville current prevalence estimate obtained using the randomized response technique is 11 percent—higher than most of the survey-based cumulative estimates. Adjusting this estimate for the fact that only 28.7 percent of all taxpayers itemize deductions, the estimate is comparable to the prevalence estimates obtained from TCMP as a frac-

58

Paying Taxes

tion of all itemizers." This suggests that the randomized response technique was relatively successful in eliciting admissions of the overstatement of deductions in one study. While Habib (1980) obtained a conflicting result, the technique may have potential for future taxpayer compliance research, especially since it can be used as the dependent variable in multivariate analyses (see Fox and Tracy, 1986). In addition, the fact that those current prevalence estimates are no lower than the cumulative estimates obtained by other researchers who generally used direct questions suggests that the usual survey-based estimates obtained using direct questions may substantially understate cumulative noncomplier prevalence; that is, respondents appear to answer in a similar manner whether a one-year or multiple-year reference period is used. Chapter 5 contains recommendations for a program to improve both TCMP and surveys as compliance measurement techniques. OTHER M E A S U R E S . Besides tax gap and prevalence estimates, TCMP data can be used to compute other statistics that measure compliance with tax reporting requirements. Two of them, the voluntary compliance level (VCL) and voluntary reporting percentages (VRPs), measure compliance in relative dollar terms—the ratio of amount actually reported by taxpayers to the amount that should have been reported. In the TCMP data base, amounts that should have been reported reflect auditors' assessments and are not adjusted to reflect appeals. V C L measures the amount of tax liability reported by taxpayers "voluntarily" (i.e., after IRS computerized arithmetic error correction but before any other enforcement contact) as a percentage of corrected tax liability as determined by TCMP examination.12 According to Fratanduono (1986), V C L for the 1982 cycle was 91.8 percent; that figure represents at least the temporary interruption of a steady downward trend in V C L , from 94.3 percent in 1965 to 91.0 percent in 1979. In contrast to the prevalence statistics indicating that noncompliance is widespread among taxpayers, the high levels of V C L reflect a high compliance level as a percentage of taxes owed. The contrast reflects the fact that most cases of noncompliance involve rather small dollar amounts. Fratanduono's estimates are not adjusted for unreported income that is not discovered by TCMP auditors. V R P is analogous to V C L but measures compliance separately for each return item: V R P is the ratio of the amount reported by taxpayers to the correct amount as determined by the auditor. For income items, VRPs below 100 percent reflect underreports of tax liability. For subtraction items, VRPs that exceed 100 percent (i.e., overstated adjustments and

Paying Taxes

59

deductions) reflect underreports of tax liability. Special tabulations from the 1982 TCMP survey were made available to the panel by the 1RS Research Division. These were used to construct Table 6, which displays compliance patterns by return item, in terms of VRPs, prevalence rates, and other compliance measures that highlight different aspects of taxpayer compliance. As Table 6 shows, VRPs vary widely across return items in patterns that suggest some conjectures about causation. For example, compliance is generally high according to the V R P measure for income items that are largely computed for the taxpayer and that are subject to withholding or information reporting, such as salaries, pensions, and interest income." V R P is low for items that are not subject to withholding and for items that require interpretations of complex regulations—tips, proprietors' income, and other (Schedule E) income are examples. Among subtraction items, taxpayers fail to take full advantage of only two—the investment credit and the adjustment for interest penalties for early withdrawal of certificates of deposit. Conventional wisdom within the 1RS is that many taxpayers are unaware of the interest penalty adjustment, and that the investment tax credit is not used by many eligible taxpayers who fail to report the income from small, low-visibility businesses. Compliance is high (i.e., V R P only slightly exceeds 100 percent) for deduction items such as I R A contributions and state and local taxes; compliance is much poorer for items that require judgments as to allowability (e.g., political contributions), imputations of amounts (e.g., casualty losses), or extensive documentation (e.g., employee business expenses). As an alternative compliance measure, Table 6 also contains complier prevalence estimates by return item from the 1982 TCMP survey. While 95 percent of all taxpayers correctly reported their wages and salaries (most of which are subject to withholding), only 70 percent reported all dividends (which are subject only to information reporting), and only 47 percent reported all tips. Easy access to records also seems to encourage compliance, which is better (using both V R P and prevalence) for mortgage interest paid to financial institutions than for mortgage interest paid to individuals. Complier prevalence rates of course increase as the underreporting tolerance is increased. But even when tax increases as large as $1,000 are ignored, complier prevalence is low for items such as proprietor's income (65 percent) and farm deductions (74 percent). Table 6 also displays the percentage of overreporting, the fraction of taxpayers who misreport items in ways that overstate tax liability.14 Rea-

— < N S. OsO 00o —I ^ N-H iCOi tX 0 and 0. Here cosh is the hyperbolic cosine function: cosh(x) = 'A (f + e~x). The marginal distribution of U\ is half-normal; u\ is distributed as the absolute value o{N(0,a\2). Similarly u2 is half-normal. The conditional distribution of u2 given U\ is folded normal, the distribution of the absolute value ofN(|p|«i, 1 - p 2 ). This implies (6)

£(»21 Ml) = \PWI +2[pV(1 - p 2 ) H

where is the standard normal cumulative distribution function. If we let the (bivariate normal) density of (vur2) be JJ(V],V2), we can make the transformation €1 = Vi + «1, e2 = v2 - «2 to obtain the density of (€i,e 2 ) as

(7)

¿(ei,e2) = I

I £f(vi,v2) f{ei - vuv2 - e2)dvidv2

= f0 jQ £t(e 1

_

2

fi

+ u2)f[uuu2)duidu2.

Statistical Issues

315

This yields a likelihood function

(8)

i= 1

L-Ylklyu-Xiifiuyx-Xih),

which is a direct generalization of equation (3) above. It could be maximized numerically to obtain the maximum likelihood estimates (MLEs) of 0 i , 02, and the parameters in the distributions of v and u. I have not attempted the integrals in equation (7), but I believe that they are tractable, in the sense that they can be reduced to a closed-form expression times the bivariate normal cumulative distribution function. A more fundamental concern is whether this approach is likely to be fruitful. Its costs are clear. First, it will lead to very cumbersome and cosdy calculations. Second, the consistency of the MLEs will (presumably) hinge on the correctness of the distributional assumptions. The gain to be had from incurring these costs is less clear. First, the MLEs will be asymptotically at least as efficient as the generalized least squares estimates. The extent of this efficiency gain is unknown, and is probably worth investigating, but it was small in the univariate case. Second, if we are interested in extracting an estimate of U\ (and/or «2), we can use E(u\ | ei,e 2 ), which should have less variability than the estimate from the univariate model, £ ( | « i | e i ) . However, while this possibly may be of some importance in the efficiency measurement exercise that motivates the production frontier literature, it is perhaps not of importance in the tax compliance setting. It is difficult to give a conclusion to this subsection without seeing some empirical results. I would like to see a bivariate system such as equation (4) estimated, both by generalized least squares and by the MLE (using the likelihood (8), for example). My suspicion is that the MLEs will not be worth the bother, while joint estimation by generalized least squares will probably be worthwhile. T w o ASYMMETRIC ERRORS

We return now to the univariate model (1), with the dependent variable being the audit-induced tax change. The two-part error term contains a symmetric error v and a nonnegative error u, with the motivation that u represents deliberate noncompliance while v represents accidental noncompliance as well as general statistical noise. A possible criticism of this error specification is that accidental noncompliance is not distributed symmetrically. For example, it is well known that even mathematical errors and other apparent mistakes tend to be predomi-

316

Appendix A

nantly in the taxpayer's favor. This could lead to a model in which v has an asymmetric distribution rather than a normal (symmetric) distribution. If we are willing to specify a particular form for the distribution of v, a likelihood function may be constructed as in equations (2) and (3), and the model may be estimated by maximum likelihood. This should be feasible, though I have some doubts as to how well we can expect to separate v from u in such a model. Even apart from questions of feasibility, however, I am not impressed by the usefulness of such an approach. The fact that mathematical errors or other mistakes are predominantly in the taxpayer's favor can simply be taken as evidence that some of them are deliberate. While this is an arguable point, certainly in terms of fitting the data the question of whether a mistake was deliberate is not a very fruitful one to argue. The dependent variable (tax change under audit) has a skewed distribution, and the component 0 is a statistical accommodation to that empirical fact. While it is nicely motivated by the distinction between accidental and deliberate noncompliance, one need not accept the distinction to decide (on the basis of data) whether the model is empirically useful and reasonable. REMARKS

In this section I have given a survey of the stochastic frontier model and discussed some possible extensions of the model that might be relevant for tax compliance research. This was done from the perspective of a possible application in which the dependent variable is the tax change induced by a TCMP audit. Such changes reflect at least in part the discovery of deliberate noncompliance. Thus the stochastic frontier model, which contains a positive error component, seems attractive. I have reached fairly pessimistic conclusions concerning the direct applicability of this model, however. The strongest motivation for the frontier model in the production function context—namely, measurement of inefficiency in individual firms—does not appear to be applicable in the tax compliance context. The main parameters of interest can be estimated consistently by ordinary least squares, without making such strong assumptions about the properties of the error terms as are required for maximum likelihood estimation of the stochastic frontier model. Furthermore, the efficiency gains obtained by making these assumptions are small. However, these conclusions depend strongly on the simplicity of the model. The linear regression model with a well-behaved additive error, as we have so far considered, cannot account for an important empirical feature of the dependent variable, namely its clustering at zero. As I discuss

Statistical Issues

317

in the next section, a successful accommodation of this feature of the data will necessarily hinge on the correctness of the distributional assumptions for certain error terms. Thus a more complicated version of the stochastic frontier model may yet be plausible and useful.

Clustering at Zero In the last section we considered models in which the dependent variable is the tax change induced by a T C M P audit. However, we ignored an important feature of the data, namely, that this variable is often exactly equal to zero, since many audits do not change the tax owed. Such a variable may be said to be clustered at zero. Clearly other potential audit-based dependent variables (e.g., change in taxable income or change in deductions) will also be clustered at zero. In this section I discuss the serious statistical problems that clustering at zero causes. I survey some standard statistical models designed to handle this problem and suggest some modifications of these models. Finally, because one reason for clustering at zero is measurement error (i.e., there is additional tax owed but the audit does not reveal it), I also discuss measurement error and the extent to which it can be incorporated into these models. STATISTICAL IMPLICATIONS OF CLUSTERING

Consider a linear regression model

(9)

yi = Xj'fi + tj,

i=l,...,N,

with the notation basically the same as in equation (1). That is, for individual i, yi is the audit-induced tax change, X, is a vector of explanatory variables, and e, is an error term. Such a model is often estimated by ordinary least squares (OLS), and the OLS estimates have many desirable properties if the errors (e,) in turn satisfy certain assumptions. In particular, if the errors are independently and identically distributed (iid) and independent of the explanatory variables (X,), then the OLS estimates are unbiased and consistent, regardless of the actual distribution of the errors. Now suppose that the dependent variable is often exacdy equal to zero. Such a dependent variable cannot be represented by a linear model such as (9), with iid errors that are independent of the explanatory variables. To see this, note thatjy, = 0 is equivalent to e, = -X,'/3. Therefore the distribution

318

Appendix A

ofyi has nonzero mass at jy, = 0 if the distribution of e, has nonzero mass at e, = —X,'¡3. Since this point varies over observations, the distribution of e must vary over observations; the e, cannot be identically distributed. Worse yet, the distribution of e, depends on since the point at which it has nonzero mass is —Xi'f3. Therefore we should expect e, and X, to be correlated. This implies that the OLS estimates will be biased and inconsistent, and that inferences based on them will be invalid. Thus application of OLS to a model in which the dependent variable is clustered at zero (or elsewhere) is not advisable. An obvious question that this raises is whether we can simply discard the observations for which y = 0 and estimate the model by OLS applied to the nonzero observations. This is actually a fairly complicated question. Its answer hinges on whether, conditional on e, —X / f i , the e, are iid and independent ofX{. If so, OLS is justified. If not, OLS will be in general be biased and inconsistent, and we have an example of what Heckman (1976, 1979) termed sample selection bias. To be able to say whether sample selection bias would be present in an analysis of the nonzero observations only, we need to make additional assumptions. Essentially, we need to expand the model to account for the occurrence of e, = —Xt'f3 with nonzero probability. This is discussed in some detail in the next subsection. For now I simply note the following, at an intuitive level. An audit reveals no tax change either because of a measurement error in the audit outcome (there should have been a tax change, but the auditor did not discover it) or because there was no tax change to be found (in which case we presume, or hope, that measurement error does not occur). If for the moment we assume away the second possibility, the question is simply whether returns without such measurement errors tend on average to have different es than returns with such measurement errors. In my opinion the answer must be yes, since on average returns with higher es have more tax change to be discovered, and this must imply a lower probability of its being undiscovered. If this is so, we should expect sample selection bias to be present if we attempt a simple analysis of the data with the no-change observations deleted. EXISTING MODELS FOR CLUSTERING

In this subsection I discuss briefly three existing models to handle clustering of the dependent variable. A less brief treatment can be found in Schmidt and Witte (1984, Chapter 4), while an encyclopedic source is Maddala (1983).

Statistical Issues

319

The first model is the so-called Tobit model, which dates back to Tobin Consider a regression model,

(1958).

(10)

where the e, are iid N(O,«?-2). However, yf is unobserved and instead we observe (11)

(We also observe X, for all i.) The Tobit model is also referred to as a censored normal regression model, because the distribution ofy* given X is normal and because the observable^ corresponds in statistical terminology to a censoring of the unobservable y* at zero. Clearly y, > 0, and y, = 0 with nonzero probability. Thus the model is often applied in cases in which the dependent variable of interest is nonnegative and zero is a common value. Clotfelter (1983) has used the Tobit model to analyze the determinants of the audit-induced change in adjusted gross income in a sample of 1969 TCMP audits. Similarly, Westat, Inc. (1980) analyzed the revenue yields of a sample of 1973 TCMP audits of business returns greater than $ 3 0 , 0 0 0 . The Tobit model is usually estimated by maximum likelihood. This requires a numerical maximization of the likelihood function, but such a maximization is not unduly hard because the likelihood function is reasonably easy to evaluate and because it is known to have only one local maximum. The maximum likelihood estimates are consistent and asymptotically efficient, so long of course as the model is correctly specified. However, the consistency of the MLEs hinges on the correctness of the normality assumption for the errors. Alternative estimators have been developed recently that are less efficient than the MLEs when the normality assumption is correct, but that remain consistent under alternative distributions of the errors, so long as the errors remain iid. Examples are Powell (1984, 1986), Manski (1985), and Fernandez (1986). In my opinion such robust estimators may be very worthwhile in the present context. There may be good reason to expect skewness in the error distribution (as discussed in detail in the previous section), so that it is unattractive to depend strongly on the correctness of an assumption of normality. A considerable disadvantage of the Tobit model in the present context is that the audit-induced changes in tax owed (or adjusted gross income, or

J20

Appendix A

other line item) are in fact not one-sided. Positive and zero changes predominate, but there are negative changes as well; the Tobit model cannot well accommodate this feature of the data. Thus Clotfelter (1983) is forced to set negative changes in adjusted gross income equal to zero. Similarly Westat, Inc. (1980) reports that they "limit the dependent variable to nonnegative values," which I find ambiguous. (It could mean that negative values are set to zero, or it could mean that such observations are discarded, and neither option is attractive.) Presumably overstatements of tax liability represent mistakes on the part of the taxpayer, but it is not appropriate to trim such statistical noise from one tail of the distribution without somehow trimming it from the other tail as well. It is more appropriate to simply regard negative values of the dependent variable as an indication that the Tobit model is incorrect. A second possible model is the two-part Tobit model of Cragg (1971). This is a generalization of the Tobit model, in the following sense. In the Tobit model one set of parameters (/^cr2) determines both the probability that y = 0 and the form of the distribution of the positive values of y. In Cragg's model there are two separate sets of parameters to determine these two features of the distribution of y. Thus the probability that y, = 0 is specified as (12)

P(y, =

=

where $ is the standard normal cdf; this is just a probit specification. Second, the density o f ^ conditional o n Z ; and on yt > 0, is assumed to be N(X,'truncated from below at zero. Thus the positive values of_y follow a truncated normal distribution. If /3i = /32 / 0. The classic example in economics is the determination of wage rates for women. Here the second equation in (13) is a probit equation explaining whether a woman works ( j a > 0) or does not work (y,2 ^ 0). The dependent variable (_y,i) of the first equation is the woman's wage, and this is observed only if she works. However, it is crucial to observe that the variable yn is assumed to be meaningful for all observations, whether or not it is observed. Thus^a is interpreted as the potential wage that the individual would earn if she chose to work, and the point of the model is to be able to estimate the effect of explanatory variables on the potential wage, without incurring the bias that may result by using only the observations on individuals that work. As a matter of notation, let crn = Cov(e,1,6,2). Then OLS ofya on X„ using only the observations for which ya is observed, will be unbiased if a 12 = 0; in this case the rule governing observability is independent of the process generating the observed ya- However, if cr12 £ 0 , OLS will generally be biased, and we call this phenomenon sample selection bias. One way to see the problem is that the regression function over all potential observation is (14)

E{ylX\X,) =X,'ßl

while the regression function over observations such that (15)

E(y,,\X„ y,2 > 0) = X',ß +

is observed is

K-

In (15), o"22 = var(e,-2), X, = 4>,/ (1 - 4>,), and , and , are respectively the standard normal density and cdf, evaluated at -X-/3 2 / Vcr^- Unless

322

Appendix A

a 12 = 0, equations (14) and (15) are different and OLS applied only to the complete observations will not give unbiased estimates of /3i. The model (13) may be estimated by MLE or by a simpler two-step method. In either case the consistency of the estimates hinges on the correctness of the assumption of bivariate normality of the errors. Alternative distributional assumptions are possible, but consistency will then in turn hinge on these assumptions. That is, no distribution-free estimation procedure currendy exists for this model. The sample selection model (13) may be applied to the present problem (analysis of determinants of the audit-induced tax change) in a fairly straightforward way, if we assume that a tax change of zero corresponds to nonobservability of the dependent variable. That is, we assume y,\ = 0 if ya < 0; otherwise^,! is determined according to (13A). H o w sensible this is seems to me to depend on whether there is real clustering at zero or whether all observed clustering is just measurement error. The model as just described does not accommodate real clustering (i.e., actual change in tax due equals zero) very well, since equation (13A) is assumed meaningful for all observations and will not generate a cluster at zero. It does handle well the case in which all observations of zero are measurement error, in the sense of failure of the audit to discover the true change in tax due. In the present context, I would expect cri2 > 0, since larger e,i mean more tax change to be discovered and thus a lower probability of observing yt\ = 0 (i.e., a higher probability o f y a > 0). Thus I believe that an analysis of only the observations with nonzero tax changes will indeed suffer from sample selection bias, and some model similar in spirit to that in equation (13) may be needed to avoid this bias. I discuss possible modifications of the model in (13) in the next subsection. A final point to note is that, of the three models discussed so far, only the sample selection model is consistent with nonzero observations that are both positive and negative. This is another advantage of the sample selection model, in the present context. While audit-induced tax changes are usually positive, they can be negative, and the Tobit model cannot handle this possibility reasonably. EXTENSIONS OF EXISTING MODELS

In this subsection I suggest some modifications of the models presented in the last subsection. These modifications will attempt to address particular issues that have already been discussed, such as measurement error and possible skewness of the errors in the compliance relationship. In particular I try to address three important questions. First, can we accommodate the

Statistical Issues

323

stochastic frontier error structure discussed in the previous section? Second, what kinds of measurement error can we handle? Third, can we distinguish real clustering of tax changes at zero from clustering caused by measurement error? In order to address these questions I discuss four possible models. The following notation is useful. Let us distinguish three (possibly different) tax amounts: T R E P , the tax reported on the return as being due; R A U D , the tax due as of the end of the audit, and T j r u e , the true tax due. The first two are observable, while the third is not. We then define Vi* = Ttrue - TREP, (16)

y1

=

A

= TAUD ~ T t r u e -

r A U D - Trep,

Note that^i* measures compliance, so that its determination is of interest. However, y\* is unobservable. We observe y\, the audit-induced tax change. Clearly y 1 = yi* + A, where A is the measurement error (of tax due) in the audit. We also note that^i will be clustered at zero, and that y\ may equal zero either becausey\* = 0 (in which case we assumed = 0) or because of measurement error, that is, A = —y\*. In the last subsection it was argued that it is reasonable to condition onjKi* £ 0 , but in the presence of measurement error it is not reasonable to condition onjyi £0. The first and simplest model to be considered is a modification of the sample selection model (13) to accommodate the stochastic frontier error structure. For now, we assume that there is no real clustering at zero. We also assume that there is no measurement error, except that some nonzero changes in tax due are undiscovered in the audit. (Thus71 = y\* orjyi = 0, and^i = 0 is the result of measurement error.) Then the model (13) can be modified slightly: (17A)

y,\ = X-ß i + v,i + Uj,

(17B)

Ja = x/ß2 + v,2,

As in equation (13), we assume (p,i,vq) iid as bivariate normal. As in equation (1), we assume «, > 0 to represent deliberate noncompliance, and a particular distribution (such as half-normal) is assumed for «,-. Furthermore u, is assumed independent of (v,i,j>,2). The independence of u, and va is a very unattractive assumption, since higher deliberate noncompliance should increase the probability of detection, but at present I don't know of a reasonable way to relax this assumption. As in equation (13), we observe

324

Appendix A

ya = 0 when y,2 s 0, and we observe yt\ as in (17A) when yl2 > 0. W e do not o b s e r v e ^ itself, but from the value ofytl

we observe whether ya — 0

or yi2 > 0. The second model to be considered is the same as (17) except that

is

added as an explanatory variable in the y!2 equation. This yields

(ISA)

ya

= Xi'fa + va + Uiu

(18B)

ya

= X-fi2 + yyiX + vi2.

This seems reasonable to me because the higher the level of noncompliance (jy,i), the more likely it is to be discovered (i.e., the higher should b e y a ) . Thus I would expect y > 0. This is a simultaneous equation model of the type commonly considered in econometrics. In the usual textbook case, equation (18B) would be underidentified without further assumptions. Possible identifying restrictions would be that one or more elements of fi2 equal zero (i.e., some variable or variables that affect compliance do not affect the probability of detection of noncompliance), or the independence of the errors in (18B) from those in (18A). However, the presence of u, in (18A) makes the model identified even without such restrictions. The above two models show that we can handle the skewness of the error distribution caused by deliberate noncompliance in a straightforward manner. W e now turn to the consideration of measurement error. In the above models, noncompliance is either undetected, or it is detected exactly, which is unreasonable since noncompliance may be detected but with error. In the third model, we relax this assumption by allowing additive measurement error. Thus we have

(19A)

y$

= Xi'fa + vn + u„

(19B)

ya

=X/p2

+ yyn*+Vi2,

(19C)

Alternatively, we can rewrite (19C) as just ya = y,f + At, where Ai = —yfi when ya — 0. Clearly, A , is interpreted as the measurement error component o f y n , just as in equation (16). W e make the same assumptions about the distribution of v,i,vi2, and u, as above, and also as above we assume that only yn and X, are observable. However, we still need to make assumptions about the measurement error

Statistical Issues

325

Ai. I will assume that the Ai are iid according to some specified distribution (e.g., normal), and thatvl, is independent of (X,, vth vi2, «,). Independence is a rather unpalatable assumption here, since measurement error should on average be larger when noncompliance is larger, but again a reasonable way to relax this assumption is not clear. I am not sure what a reasonable distribution for Ai would be, but normality is certainly unreasonable, since Ai < 0 must be much more common than A, > 0. I would probably assume Ai < 0 and therefore assume a one-sided error distribution for it. As with the one-sided error «„ there is frankly not much guidance for its choice. It is not clear whether this model is identified, in the sense that we can hope to distinguish the measurement error Ai from the other errors (v,i, p,2, Ui). If A, were normal, we could not do so since A, would be indistinguishable from Vii. However, if At has a distribution unlike those of the other errors, its distribution is probably identified. In any case, the parameters of interest (/3i, and secondarily fii) remain identified even with measurement error on y*a. It is also worth noting explicitly that measurement error onX, is far more serious than measurement error onjy*j. It can be handled only by making assumptions even less realistic than those made so far (e.g., known variance of the measurement error). This may be a serious problem if some of the explanatory variables are other line items on the return for which noncompliance is also likely. The final and most complicated model I consider is designed to allow real clustering of tax changes at zero, as well as measurement error. It is a combination of Cragg's model and our modification of Heckman's sample selection model. It is of the form (20A)

y.0 = X/p0 + vi0,

(20B)

y.* = Xj'Pi + va + «„

(20C)

y, 2 = x;p2 + yyn* + vl2,

(20D) We make the same assumptions as above, and in addition assume v,o to be iid as standard normal and to be independent of (v,i, vl2, u„ A,). We observe only y,\ and X,. Equation (20A) accounts for real clustering. The intended interpretation

326

Appendix A

is that yt\* = 0 (and yn = 0) if yi0 < 0. The probability of this event is (-Xi'/3o), which is consistent with equation (12) in our discussion of Cragg's model. The interpretation of (20B) is that it gives the distribution of yft conditional on yfi^O (yi0 > 0). Its interpretation as a conditional distribution justifies the assumed independence of (va, u t ) and v,o. As before, equation (20C) is interpreted as determining the probability that yn = 0 when jy.f^O (i.e., because of measurement error). Finally, (20D) governs measurement error in the observations of the nonzero observed tax changes. Amazingly enough, I believe that this model is identified. That is, we can indeed separate real clustering at zero from clustering due to measurement error. To see how the model accomplishes this, note that (21)

P(ya = 0) = P(ya

0 or_y,0 < 0)

= 1 - P{yi2 > 0 and_yl0 > 0) = 1 - P(yi2 > 0)P(yio > 0),

using independence of vi0 from the other errors. Given that P(yn = 0), P(ya > 0) and P(yt0 > 0) all depend on X„ the model imposes enough structure on the way they change with X, to identify /30 and J82 separately. (It could not do so, obviously, if P(ya > 0) and P(yi0 > 0) were constant over i.) A similar model with a similar conclusion is given by Poirier (1980). As in the Poirier model, however, identification is accomplished by having assumed specific distributions for the errors, specific (linear) functional forms for the deterministic portions of equation (20), and independence of virtually all error terms. This is unfortunate, but should have been expected. Some tax changes are observed to be zero, and basically without further information we are supposed to determine whether this is because the true tax change is zero or because a true nonzero tax change has not been detected. Such a distinction is unlikely to be made without paying a high price in terms of assumptions required. Estimation of any of the models of this subsection will require the numerical maximization of a likelihood function. Especially for the more complicated models, this is likely to involve rather difficult and extensive computations. Furthermore, the consistency of the MLEs will depend on the correctness of the specification of the model (including the distributional assumptions for the errors), and this is unfortunate in view of the long list of assumptions underlying these models.

Statistical Issues

327

REMARKS

The models required to handle clustering of the dependent variable (audit-induced tax change) at zero are complicated. This is unfortunate because such clustering is inconsistent with the assumptions of simpler models, such as the linear regression model, and because it is not legitimate (except under implausible assumptions) to simply delete observations for which the dependent variable is zero. However, it is possible (under admittedly strong assumptions) to devise models that account for clustering, which include an accommodation of measurement error, and indeed which can separate clustering due to measurement error from genuine clustering of the dependent variable. Given the difficulties caused by clustering, it may seem attractive to use data that are not clustered. For example, the 1969 and 1979 T C M P data are available, aggregated to the three-digit zip code level, along with other relevant demographic data at the same level. (Such aggregation is obviously one way in which the I R S may avoid confidentiality problems in a publicuse version of the data.) I will presume that these data are the averages of the corresponding individual data, with the average taken over individuals in the T C M P file who reside in the three-digit zip code area. These data are presumably not clustered at zero, since, for example, it would be unlikely that no one in the T C M P file for a given three-digit area had a nonzero tax change. Thus it is tempting to analyze them with a simple model like the regression model. Now, it is obvious that if a linear regression model is appropriate at the individual level, it will be appropriate at the aggregate level; averaging over observations preserves linearity of the model. However, as argued above, clustering at zero in the individual data implies that a linear regression model is inappropriate at the individual level. That being the case, there is really no good justification for a linear regression model at the aggregate level. I do not at this point have a better alternative to suggest, but perhaps one can be found. A final remark is that my pessimistic conclusions at the end of the previous section concerning the usefulness of stochastic frontier models do not apply to the more complicated models of this section. In these more complicated models the consistency of one's estimates hinges on the correctness of the distributional assumptions (which is not so in a linear regression model). If the stochastic frontier error structure seems reasonable in the compliance equation, as it does to me, it should be used.

328

Appendix A

Panel Data By panel data I mean data on each of N individuals for T time periods (years, in the present context). This is stronger than assuming a crosssection of size N for each of T years, since in panel data the same individuals are followed over time. For data derived from TCMP audits, it is clear that the panel length T would have to be very short. Repeated auditing of an individual, year after year, seems sure to change his or her behavior, so that observations past the first few for any individual would be of questionable use. (Of course, if we have panel data we can test whether auditing an individual affects his or her behavior, and the desire to conduct such a test is one possible motivation for collecting panel data.) This retesting bias could be a serious problem even in a very short panel (e.g., T = 3). However, a short panel free of retesting bias could presumably be constructed by simultaneously auditing the returns of the three most recent tax years, for example, for a set of N individuals. The maximum possible length of such a panel would be determined by the statute of limitations for tax audits. Alternatively, we could analyze cross-sections for each of T years, where T could be as large as the number of years that TCMP audits have been done, if we do not require that the same individuals be in each crosssection. We could even match up individual characteristics so that each cross-section contains individuals with (more or less) the same values for potential explanatory variables. Thus an important question to be answered is for what purposes is it necessary to follow the same individuals over time. The existing literature on panel data discusses three motivations for its use: (i) to increase efficiency of estimation by using more observations; (2) to control for potential biases caused by unobserved individual characteristics; and (3) to study the dynamics of the process being analyzed. I discuss these three possibilities in turn. INCREASED EFFICIENCY OF ESTIMATION

One possible motivation for the use of panel data is to increase the efficiency of estimation. At the simplest level, the sample size may be increased by using more than one cross-section, thus leading to more efficient estimation than would be possible with a single cross-section. For example, if we were interested in estimating a production function using annual data at the state level, we might worry that thefiftyobservations in a single year's data would not be enough to yield precise estimates. However,

Statistical Issues

329

if we had twenty years' data on each of the fifty states our sample size would be 1,000, and we would expect better results with 1,000 observations than with 50. This argument may not be very compelling with T C M P data, because the sample size in a single year's data is already very large. It may perhaps be relevant in aggregated versions of the T C M P data. A more subtle reason why panel data may increase the efficiency of estimation is that some explanatory variables may vary insufficiendy in the cross-sectional dimension but more in the time-series dimension (or conversely). The classic example in economics is a budget study that attempts to explain individual consumption patterns as a function of prices and income. Income varies considerably in a cross-section of individuals, but prices do not. (There is some regional variation in prices, but not much.) Therefore a cross-sectional study may estimate the effect of income on consumption precisely, but not the effects of prices on consumption. Conversely, prices vary over time but income varies less over time than over cross-sections (especially after the change in nominal income due just to changes in the general price level is removed). Therefore a time-series study may estimate the effects of prices on consumption precisely, but it may do a poor job of estimating the effect of income changes. An obvious solution is to use panel data so that both cross-sectional and temporal variation in explanatory variables may be exploited. Similar cases may exist in the present context. For example, it is clear in the T C M P cross-section that compliance levels are lower for younger taxpayers than for older taxpayers (e.g., Clotfelter 1983, Table 2). It is not clear whether this is an effect of age or of generation, and the distinction is very important in terms of its implications for future compliance levels. If it is an effect of age, we would expect the noncompliant younger generation to become compliant as it gets older, and the general level of compliance should not change much. However, if the noncompliant younger generation is fundamentally different from the older generation (because attitudes toward the I R S have changed over time, for example), it may remain noncompliant as it ages, and the general level of compliance will fall as the last remaining compliant generation dies off. Because age and generation do not vary independently in the cross-section, this distinction cannot be made in a simple cross-section. However, generation is constant over time while age changes, so we can get evidence about this distinction by pooling cross-sections taken at different points in time. Given the practical difficulties in obtaining panel T C M P data, it is very

330

Appendix A

important to note that nothing in the above discussion requires that the same individuals be followed over time. For example, to separate age and generation effects we need to pool cross-sections over time, but we do not really need panel data. In certain contexts (e.g., budget studies) it is cheaper to collect panel data than to collect independent cross-sections in different time periods, since the effort put into the selection and location of individuals need not be repeated every year. In the case o f T C M P data, however, the retesting bias problem is potentially very severe, and I would not expect any panel data set to be long enough (cover a large enough number o f years) to be satisfactory. A series o f cross-sections taken at different points in time, with some care to ensure comparability of data definitions and sampling schemes, would be more feasible and more useful in increasing the efficiency o f estimation. CONTROLLING BIAS D U E TO UNOBSERVABLES

Another motivation for using panel data is that it may enable elimination from the model o f bias caused by the omission o f unobservable individual characteristics. The classic example in economics is the estimation o f a production function for a cross-section o f farms. Suppose that for each farm we observe output and several inputs, such as labor, capital, acreage, and fertilizer. Suppose we do not observe soil quality. Clearly soil quality affects output, so it is implicitly part o f the error term in the regression model (production function) we estimate. It is also clear that soil quality and input usage should be correlated; for example, a farmer may use less fertilizer on good soil than on poor soil. Thus a regression o f output on inputs, not correcting for differences in soil quality across farms, will yield biased estimates. For example, the coefficient o f fertilizer will presumably be biased toward zero, since we do not take into account that fertilizer is applied disproportionately to poor soil. This potential bias may be avoided, without observing soil quality, if we have panel data and if we are willing to assume that the effect of soil quality on output is time-invariant. That is, we assume a model o f the form (22)

ylt = X,t'fi

+ a , + e, ()

i = 1,. . .,N,

t = 1,. . ,,T,

where i = 1,. . ., N indexes farms; t = 1,. . ,,T indexes time periods (years); y is output and X is a vector o f inputs (probably measured in logarithms, in this example); a, is the intercept for farm i; and e is a nicely behaved error term that is independent o f a a n d X Thus each farm has the same slope coefficients (/3) but a different intercept (a,). I f the effect o f soil

Statistical Issues

331

quality on output is additive and time-invariant, it is absorbed into the farm-specific intercept. We can obtain unbiased estimates of f3 by estimating equation (22) by least squares, treating the a, as parameters. This can be done by using dummy (indicator) variables for the individual farms. It is equivalent to least squares of y on X after transforming the data to deviations from individual-farm means, that is, regression of (y„ - y,) on (Xa - Xi). Clearly the idea is that differences in soil quality (or any other time-invariant variables) disappear when we transform to deviations from individual means. A second example, taken from the field of labor economics, concerns estimations of the effects on an individual's wage rate of characteristics such as age, race, sex, education, and occupation. If individuals differ in ability, which affects the wage and which is correlated (we hope) with education, then a regression of wage on individual characteristics will yield a biased estimate of the coefficient of education. A regression using data in deviations from individual means will yield unbiased results if the effect of ability on wage is time-invariant. In the present context, it is possible to imagine that individuals differ in an unobservable characteristic that I will call honesty, which affects compliance behavior. If honesty is correlated with the explanatory variables in the compliance relationship, then the results from the analysis of a single cross-section will be biased. This bias can be avoided in panel data so long as honesty is time-invariant. Under the assumptions we have made, equation (22) would be called a fixed-effects model since the individual effects a, are treated as fixed parameters. The transformation to deviations from individual means is called the "within" transformation, since it corresponds to the within-class variation in an analysis of covariance. Least squares after the within transformation is often called the "within" estimator. In fact, it is just the usual analysis of covariance estimator. A disadvantage of this fixed-effects treatment is that one cannot use time-invariant explanatory variables, since they vanish under the within transformation. Many demographic variables (e.g., race, sex) are timeinvariant. To be able to estimate the effects of time-invariant regressors, it is necessary to be willing to assume (on an a priori basis) that some of the regressors are uncorrelated with the individual effects. A detailed discussion of this issue is beyond my scope here, but briefly the number of regressors assumed uncorrelated with the individual effects must be at least as large as the number of time-invariant regressors. For more details, see Hausman and Taylor (1981).

332

Appendix A

In order to eliminate the bias due to time-invariant unobservables, we do not need a long panel. We only need (at least) two observations per individual. The above discussion has considered the case of a linear regression model. In more complicated models, such as the models suggested above to handle clustering, unobservables cannot be handled in such a straightforward manner. In particular, if we let the intercept in one or more of our equations vary over individuals in an unrestricted fashion, the maximum likelihood estimates of the slope coefficients are not necessarily consistent. This is a reflection of an incidental parameters problem; as the sample size increases, when more individuals are sampled, the number of parameters also rises. This does not cause any notable problems in the linear regression model, but it may in more complicated models. For a statement of the problem and solutions for some specific models, see Chamberlain (1980, 1984,1985). C O N S I D E R A T I O N OF D Y N A M I C S

A third possible reason for using panel data is to study the dynamics of the compliance process. At the simplest level, we may be curious as to whether there is correlation over time, at the individual level. That is, does noncompliance at some point in time raise the likelihood of noncompliance in the future? To give a statistical answer to this question clearly requires that each individual be observed at two or more points in time. With panel data we can indeed hope to determine whether noncompliance is correlated over time. We can also make some more subtle but fundamental distinctions about the reasons for such correlation. Suppose for example that we sample a number of individuals in two different years, year 1 and year 2. It is straightforward to separate the sample into the groups that were compliant and noncompliant in year 1 and to compare the compliance rates of these groups in year 2 . 1 presume that the compliance rate in year 2 would be higher for the group that was compliant in year 1 than it would be for the group that was noncompliant in year 1. In that sense there would obviously be correlation over time. But it is important to realize that there are (at least) two very different possible reasons for such correlation. One possible reason is that, at the individual level, being noncompliant in year 1 has an effect on compliance behavior in year 2. Perhaps if I cheat on my taxes in year 1 , 1 discover how easy it is to do so and, having learned that it is easy, I do it again in year 2. Note that this a genuine effect on

Statistical Issues

333

individual behavior and it corresponds to correlation in behavior over time even if we condition on an exhaustive list of individual characteristics. A second possible reason is simply that individuals who have characteristics that make them more likely to cheat in year 1 will have more or less the same characteristics in year 2. If honesty is genetic, for example, people with honest genes will tend to cheat in neither year, while people with dishonest genes will tend to cheat in both years. This will generate correlation of observed outcomes (noncompliance) over time, but in this case there is no correlation over time in the distribution of compliance conditional on an exhaustive list of individual characteristics (including some, like genes, that are surely unobservable). The first possibility is called state dependence, because the state (compliance or noncompliance) occupied in year 1 influences the state occupied in year 2 for any individual. The second possibility is called heterogeneity, for obvious reasons. Given heterogeneity, the individual's state in year 1 helps to predict his or her state in year 2 only because it conveys information about the individual's characteristics. The distinction between state dependence and heterogeneity occurs in many different kinds of models. It is an important distinction because policies designed to prevent loss of virginity are potentially valuable under state dependence but much less useful under heterogeneity. Given panel data, we can hope to distinguish between these two phenomena. The statistical techniques that are necessary depend on the model used, but a good general treatment is provided in Chamberlain (1984, 1985). The usual treatments assume a short panel of many individuals, which is exactly the kind of panel that may be feasible in the present context. ISSUES OF DATA COLLECTION AND SAMPLE D E S I G N

As discussed earlier, construction of a long T C M P panel data set is problematic because of the problem of retesting bias. A short panel could be constructed by simultaneously auditing the returns for several different years for a given individual. A short panel is obviously less informative than a longer panel would be, but it is sufficient to let us control for potential biases caused by time-invariant heterogeneity. A short panel may also be sufficient for studying the dynamics of the compliance process; presumably its sufficiency depends in part on how long the lags in the process are. In terms of increasing the efficiency of estimation, I argued earlier that a series of cross-sections (not sampling the same individuals repeatedly) may be sufficient. By sampling different individuals we can hope for a longer

334

Appendix A

data series without retesting bias. We cannot eliminate the bias due to heterogeneity nor study the dynamics of the compliance process with such data, however. An apparent possibility is to construct an artificial panel by sampling an individual in period i, sampling a different individual with identical demographic characteristics in period 2, and so on. I don't see much point in such an artificial panel, however. Insistence on carrying the same demographic characteristics through time limits the variability of the explanatory variables and is thus not recommended from the point of view of efficiency of estimation. And from the point of view of controlling bias or studying dynamics, an artificial panel is useless. REMARKS

In my opinion, it would be very worthwhile for the IRS to collect a short TCMP panel data set by simultaneously auditing several years' returns for a set of individuals. It would also be very worthwhile for them to construct a longer series of TCMP cross-sections, in which different individuals are represented over time but some care is taken to ensure comparability of data definitions and sampling strategies over time. Many interesting questions (only a few of which have been discussed in this section) could be answered using these data sets and could not be answered with any single cross-section.

Concluding Remarks In this appendix I have discussed a number of statistical issues that arise in the analysis of individual TCMP data or other similar individual-level quantitative data. Much of the discussion is also relevant to the analysis of aggregated versions of the TCMP data. Little of it, however, is relevant to an analysis of survey data, chiefly because in surveys the interesting compliance questions tend to yield data that is qualitative (e.g., a "yes" or "no" answer) rather than quantitative (e.g., an audit-induced tax change). Another difficulty with survey data is that, because the honesty of answers to sensitive questions is doubtful, unusual sampling schemes may be employed, and these lead to unusual statistical issues. As an extreme example, in the "locked-box" technique, an individual's answer to a sensitive compliance question ("Did you stretch the truth a little in order to pay fewer taxes for 1978?") is placed in a locked box. Anonymity is assumed because the answers to demographic questions are kept separate.

Statistical Issues

335

This sampling scheme unfortunately makes it impossible to estimate the effects of demographic variables on compliance, although the average level of compliance may be estimated. Another example is the so-called randomized response sampling scheme, in which the individual answers either the sensitive compliance question or an innocuous question, according to the outcome of a coin toss or similar simple random outcome. Here the response may be kept together with demographic data on the individual, since there is no way to know which question the individual answered. A statistical methodology exists for estimating the average level of compliance from randomized response data (see, e.g., Tracy and Fox, 1981), but surprisingly enough no methodology apparendy exists for estimating the effects of explanatory variables on the level of compliance. (That is, we can estimate a proportion but cannot do the equivalent of a regression.) I believe that such a methodology could be developed. To the extent that we believe that the randomized response scheme elicits honest responses, development of this methodology would seem to be an important task. One issue that was discussed in some detail is the statistical implication of the assumption that tax fraud is one-sided. (We presume that no one deliberately overstates his or her tax liability.) This implies that certain data, such as audit-induced tax changes, should be skewed in a predictable direction. I discussed a particular model, the stochastic frontier model, whose error structure can accommodate such one-sided errors. However, such an accommodation of one-sidedness is largely unnecessary in simple models like the regression model. In more complicated models, its accommodation is more important (assuming that one-sidedness exists in the process generating the data) but also more complicated. A second issue that was discussed is the clustering of observed tax changes at zero. This has profound statistical implications, since its occurrence is inconsistent with simple models like the regression model. An important corollary is that simple models are inappropriate even for aggregated data sets in which clustering no longer appears. I discussed the advantages and disadvantages of standard statistical techniques to handle clustering and proposed some complicated extensions of these models. I conclude that we can handle clustering at zero, and that we can even distinguish alternative sources of it, but only under very strong assumptions. The third issue that this appendix has discussed is the construction and use of panel data. Here my conclusions are more optimistic. While perhaps only a short panel data set (i.e., one covering only a few years per individual) may feasibly be constructed, such a data set will enable researchers to

336

Appendix A

answer a variety of questions that cannot be answered from a single crosssection or even from a series of independent cross-sections. Furthermore, for other questions for which a long panel is important, a series of independent cross-sections may suffice. Thus I am able to recommend some specific, feasible data collection strategies and to indicate the advantages which the new data would offer. I conclude by noting that I personally do not feel that the research community can be expected to make much progress in understanding the determinants of taxpayer compliance without having access to the TCMP data. To convince the IRS to release these data, it is necessary to find ways to ensure the confidentiality of individuals' identities. I presume that this is not an insuperable problem. But it is also necessary to explain the uses to which the data might be put, and the questions they might help answer. I hope this discussion is a useful part of this task.

References Aigner, D.J., Lovell, C.A.K., and Schmidt, P. 1977 Formulation and estimation of stochastic frontier production function models. Journal ofEconometrics 6:21-37. Chamberlain, G. 1980 Analysis of covariance with qualitative data. Review of Economic Studies 27:225—238. 1984 Heterogeneity, omitted variable bias and duration dependence. Chapter 1 in J. Heckman and B. Singer, eds., Longitudinal Analysis ofLabor Market Data. Cambridge: Cambridge University Press. 1985 Panel data. Chapter 22 in Z. Griliches and M. Intriligator, eds., Handbook ofEconometrics, Vol. II. New York: North Holland. Clotfelter, C.T. 1983 Tax evasion and tax rates: An analysis of individual returns. Review of Economics and Statistics 65(3) :b63—373. Cragg, J.G. 1971. Some statistical models for limited dependent variables with application to the demand for durable goods. Econometrica 39:829—844. Fernandez, L. 1986 Non-parametric maximum likelihood of censored regression models. Journal ofEconometrics 32:35-58. Hausman, I.A., and Taylor, W. 1981 Panel data and unobservable individual effects. Econometrica 49:1377-1398.

Statistical Issues

337

Heckman, J.J. 1976 The common structure of statistical models of truncation, sample selection and limited dependent variables and a simple estimator for such models. Annals ofEconomic and Social Measurement 5:475-495. 1979 Sample selection bias as a specification error. Econometrica 47:153-162. Johnson, N.L., and Kotz, S. 1972 Distribution in Statistics: Continuous Multivariate Distributions. New York: John Wiley & Sons. Jondrow, J., Lovell, C.A.K., Materov, I.S., and Schmidt, P. 1982 On the estimation of technical inefficiency in the stochastic frontier production function model. Journal ofEconometrics 23:269-274. Kmenta, J. 1971 Elements ofEconometrics. New York: Macmillan. Maddala, G.S. 1983 Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press. Manski, C.F. 1985 Semiparametric analysis of discrete response: asymptotic properties of the maximum score estimator. Journal ofEconometrics 27:313-333. Olson, J.A., Schmidt, P., and Waldman, D.A. 1980 A Monte Carlo study of estimators of stochastic frontier production functions. Journal ofEconometrics 13:67-82. Poirier, D.J. 1980 Partial observability in bivariate probit models. Journal of Econometrics 12:209-217.

Powell, J.L. 1984 Least absolute deviations estimation for the censored regression model. Journal ofEconometrics 25:303-326. 1986 Symmetrically trimmed least squares estimation for Tobit models. Econometrica 54:1435-1460. Schmidt, 1976 Schmidt, 1984

P. Econometrics. New York: Marcel Dekker. P., and Sickles, R. C. Production frontiers and panel data. Journal of Business and Economic Statistics 2:367-374. Schmidt, P., and Witte, A.D. 1984 An Economic Analysis of Crime and Justice: Theory, Methods and Applications. New York: Academic Press. Stevenson, R.E. 1980 Likelihood functions for generalized stochastic frontier estimation. Journal ofEconometrics 13:57—66. Tobin, J. 1958 Estimation of relationships for limited dependent variables. Econometrica 26:24-36.

338

Appendix A

Tracy, P.E., and Fox, J.A. 1981 The randomized response approach to criminological surveys. Chapter 3 in J.A. Fox, ed., Methods in Quantitative Criminology. New York: Academic Press. Waldman, D.M. 1984 Properties of technical efficiency estimators in the stochastic frontier model. Journal ofEconometrics 25:353—364. Westat, Inc. 1980 A Procedure for Estimating Taxpayer Response to Changes in IRS Audit Coverage. Paper prepared for the Internal Revenue Service, March 1980, by Westat, Inc., Rockville, Md.

Robert F. Boruch Appendix B: Experimental and Quasi-Experimental Designs in Taxpayer Compliance Research

To tax and to please, no more [than] to love and be wise, is not given to men Edmund Burke, On American Taxation, 1774

This appendix considers the use of randomized field experiments and certain quasi-experimental designs to assess tax compliance projects. The presumption is that readers are acquainted with the general idea of randomized field tests of social and institutional programs. It begins with a brief introductory treatment of the topic, followed by a section illustrating randomized tests in tax administration contexts. The third section focuses on feasibility of randomized tests in the tax compliance arena. The criteria for judgments about feasibility are taken from the oudine of issues developed by Riecken et al. (1974) to aid in resolving the problems engendered by experimentation. They include legal and ethical issues, scientific and statistical matters, political and institutional questions, and managerial requirements. The fourth and fifth sections consider issues involved in both experimental and quasi-experimental design. The issues are general but they demand tailored approaches in the tax administration arena. These issues include: the choice of units of analysis, estimating the reactive effects of field experiments, choosing the number of sites, understanding generalizability of the Departments of Psychology and Statistics, Northwestern University. Comments by JeffRoth, Jan Kmenta, Al Reiss, and Richard Schwartz helped greatly to improve this paper. An earlier version was presented at the Symposium on Taxpayer Compliance Research, Padre Island, Texas, January 15-17, 1986. Background research on the topic has been supported by the National Science Foundation.

34o

Appendix B

field test, and understanding how to couple experiments to quasiexperimental and survey research on compliance.

D E F I N I T I O N AND ILLUSTRATION

By "randomized field experiment" here is meant a situation in which individuals (or organizations) are randomly assigned to one of two or more regimens to understand the relative effectiveness of the regimens. The technology permits a probabilistic statement to be made about the results. For example, a sample of IRS offices might be used to understand which of two procedures reduces the cost of processing certain kinds of returns. Half the sample of offices is assigned to one regimen and half to a second regimen, perhaps the conventional (control) approach. The offices are monitored over time to determine which procedure yields the largest cost reduction without an appreciable decrease in the quality of the service provided. The main object of such an experiment is to obtain an unbiased, relatively unequivocal estimate of the relative effectiveness of the two procedures. "Relatively unequivocal" here means avoiding the problem of competing explanations—for example, that some extraneous variables, rather than the new procedure, produced the effects observed. Estimating the effects of these variables is difficult, often impossible, in normal social contexts. "Unbiased" means producing a fair estimate of the difference between the regimens. Technically, the simplest randomized test procedure depends on the use of a regression model with one explanatory variable that is controlled and a second explanatory, the error, that is uncorrelated with the controlled variable by virtue of the randomization. The absence of correlation is crucial to making interpretable estimates of parameters in the model. See, for instance, any good text in experimental design or econometrics. BACKGROUND

A major stress on using randomized field experiments to plan and evaluate social programs began in the late 1960s and early 1970s in the United States. The main scientific justification for the emphasis was development of better, less biased, less ambiguous estimates of the effects of social programs.

Experimental Designs

34i

The second, related justification depended heavily on political context. For instance, the Social Science Research Council's (SSRC) Committee on Social Experimentation, the President's Science Advisory Committee under the Nixon administration, the National Research Council, the Brookings Institution, and others contributed to the effort to understand how better evidence could be introduced into policy debates about effectiveness of new programs (see Riecken et al., 1974, for an early review). The individuals who contributed to the early efforts (see Riecken et al., 1974) include economists such as Joe Newhouse, Alice Rivlin, and Harold Watts; methodologists-social scientists such as Donald Campbell, Frederick Mosteller, William Kruskal, and Peter Rossi; and social scientists such as Henry Riecken. The newer contributors are acknowledged in the citations here. The SSRC work led to Riecken and others' (1974) state-of-the-art monograph, Social Experimentation. It treated managerial, political, institutional, ethical, legal, and scientific problems engendered by trying to do good, controlled tests of government initiatives. Since 1974 a variety of new research monographs and general texts on the topic have been produced for various disciplines. The Federal Judicial Center's (1983) Experimentation and the Law, for instance, dedicates serious attention to ethical aspects of randomized tests of new court procedures. Ferber and Hirsch (1982) and Fienberg, Singer, and Tanur (1985) are remarkable for their coverage of experiments to develop better evidence about economic projects, programs, and policy. Recendy created journals, such as Controlled Clinical Trials in medicine and health services and Evaluation Review contain material that is relevant to controlled experiments in the social sector. Randomized Experiments and Compliance-Related Research Randomized experiments designed to understand how to improve tax administration have indeed been run. They are discussed briefly here partly to illustrate and partly to identify distinctive features of the work. Randomized experiments eliciting information about sensitive topics have also been run in related areas, notably law enforcement and civil and criminal justice. They have been used in education as well. Examples are discussed, partly to capitalize on others' experience. For good bibliographies in these and related areas see Matt (1988), Dennis (1988) and Boruch, McSweeny, and Soderstrom (1978).

342

Appendix B

T A X COMPLIANCE THE ACCOUNTS RECEIVABLE TREATMENTS STUDY.

T h e I R S ' s Accounts

Receivable Treatments Study, described by Perng (1985), involves tests of five experimental methods for collecting the balance due from individual delinquent accounts. In the conventional (control) approach to collection, four standard notices are sent to delinquents with five weeks between the first and second notices, and with three weeks, four weeks, and four weeks between subsequent waves (5-3-4-4). The five new methods were variations on this basic approach. For instance, in the first variation, an installment offer was included in the third notice. The second variation included an installment offer and extended the time between second and third notices to fifteen weeks. Two other variations involved telephone calls (three of them) between mailed notices. Over 46,000 delinquent individuals were involved in the study, excluding control group members. Randomization was accomplished by using the ending digits of Social Security numbers. This process is interesting in that social security numbers provide an audit trail that is viewed as legitimate by the courts; the process is legally witnessed at the IRS, for example. It is described in the IRS's Compliance Measurement Handbook. The first lesson of the study is, of course, that the IRS has indeed run a randomized field experiment on collection methods. The administrative procedures tested might be regarded as innocuous. Still, it is a beginning and an important one. The second lesson is that special provisions for legally witnessed randomization must be used, as in draft lotteries, gaming lotteries, and so on. It is not clear that there is a need for "auditable" random numbers in other areas. Third, no unusual managerial control or political institutional problems seem to have been encountered. LEGAL SANCTIONS AND CONSCIENCE.

T h e Schwartz and Orleans (1967)

experiment is something of a classic. The object was to determine whether two survey interview processes could affect subsequent taxpayer behavior. One of the processes posed survey questions that stress the taxpayers moral obligations to comply with law. A second process stressed the legal sanctions engendered by failure to comply. A third (placebo) group was asked general survey questions on political and civic issues with emphasis on tax policy; this group constituted the control condition. Interviews were conducted a month before individuals filed their tax returns. Both threat of sanction and appeal to conscience appear to work in the sense of increasing reported income. But sample sizes were small: 89-92 in the three main groups.

Experimental Designs

343

The first interesting feature of the work is that the experiment was run by university scholars with IRS cooperation. Such collaboration, and other examples of it in the studies discussed below, is remarkable; other enforcement/ investigatory agencies do not find collaboration desirable or feasible. Second, the potential privacy problems were resolved simply. IRS provided the researchers with only marginal distributions for the experimental groups. No records on identifiable individuals were disclosed. Third, the experiment produced interesting, potentially important results. This raises the question whether the Schwartz-Orleans experiment should be and can be replicated with larger samples, perhaps paying more attention to theory, to understand whether and how the findings bear on specific tax policy regulation or law. METHODS FOR ESTIMATING NONCOMPLIANCE IN SURVEYS.

Aitken and

Bonneville's (1980) General Taxpayer Opinion Survey includes an informative randomized trial. The trial was designed to understand which of several alternative methods of questioning people about tax cheating led to more accurate estimates of the incidence of cheating; it was sponsored by the IRS Office of Planning and Research. It is compatible with concerns expressed by Witte (1987), among others, that more accurate approaches are needed to gauge nonfiling and evasive income reporting. The three methods that were tested were: • randomized response to questions about cheating; • locked-box approach to questions; and • indirect questions, e.g., comments on approaches to cheating. Over 4,800 individuals were assigned randomly to one of the three groups. The randomized response methods essentially involve the respondents' injecting a response with probabilistic error so that the interviewer cannot tell what the true state of the individual is, but statistical analysis can adjust for the controlled error in large samples (see the section on feasibility below). The locked-box approach generally requires the respondent to seal a response and place it in a locked box or mail box in the presence of the interviewer. The results were promising for the randomized response relative to the locked-box method. Perceived anonymity was higher and admission of tax cheating was consistently higher. The first implication of this work is that such randomized tests of alternative methods of estimating cheating are feasible. This extends a fine tradition of methodological work by the U.S. Census Bureau, the National Center for Health Statistics, and others. Second, the randomized response worked well in this instance. The method does not always work, however (see the review in

344

Appendix B

Boruch and Cecil, 1979). The implication is that field tests need to be done to understand when and why it works. LONG RANGE TAX FORMS SIMPLIFICATION STUDY.

The Long Range Tax

Forms Simplification Study (IRS, 1983) focused on two kinds of alternatives to the conventional 1040A. The first, i04oS-Customized, was tailored to each of four statuses of taxpayers, as they identified themselves in a previous year. The second, io4oS-Consolidated, could be used by a 1040A filer of any filing status. Both alternatives were developed under laboratory conditions prior to field testing. Both involved a consortium approach to redesign, e.g., simplifying language; changing format from two columns to one; increasing use of color, worksheets, and illustrations; and providing tax hints. Each alternative was assigned to 14,000 taxpayers, and a control of 28,000 taxpayers was constructed. The experiment, executed in Georgia, appears to have involved randomized assignment. (The ambiguity in the report is noteworthy; see Betsey Hollister, and Papageorgiou, 1985, on similar ambiguity in youth employment experiments). The work was based on a stratified random sample design constructed so as to ensure that each treatment replicate was as similar as possible to the Atlanta population of taxpayers. Use of the alternative forms was voluntary. That is, subjects could use the assigned alternative or the conventional forms. Similarly, responding to a follow-up questionnaire was voluntary. Results of analyzing the returned forms suggest that fewer arithmetic errors were engendered by the io4oS-Consolidated relative to the 1040A (6 percent versus 8 percent) and fewer errors by separate line item (12 versus 15). Taxpayers liked the new forms better by two to one. The io4oS-Customized, however, fared badly, engendering higher error rates (8 percent versus 6 percent) than the consolidated forms. This work is remarkable in several respects. It demonstrated that a laboratory approach plus field testing, with an independent consortium of private contractors, can be productive. It illustrates the feasibility of testing forms in formal experiments. It is a nice example of redesign features and the use of field test results to construct subsequent alternative 1040EZ and the 1040 with instructions. The work is instructive for illustrating two problems of randomized experiments. The response rate to the questionnaire (10 percent) is low by conventional standards. Moreover, there are small but noteworthy differences in response rates across treatment categories: 10.7 percent for the i04oS-Consolidated, 10.2 percent for the io4oS-Customized, and 7.6

Experimental Designs

345

percent for the 1040A. What adds credibility to an otherwise suspicious sample is the lab results ("consistent" with these). The second problem concerns the rate at which individuals choose to complete the alternative forms they are assigned (instead of choosing the conventional 1040A or 1040): 32 percent for the i04oS-Consolidated and 30 percent for the i04oS-Customized. This is a lower bound in the sense that the i04oS-Consolidated, for instance, is an alternative only to the 1040A. Taking this into account boosts the choice rate to 36 percent (3,707/6,608). This low choice rate is attributable partly to the fact that many respondents had their forms prepared by tax practitioners, and over 85 percent of these used the 1040A instead of the experimental forms. N o analysis of analytic biases engendered by the rate is given in the report. COMPLIANCE RELATED EXPERIMENTS

POLICE PROCEDURES. In the Minneapolis Domestic Violence experiment, the object was to understand how police should handle calls on domestic violence cases (Sherman and Berk, 1984). Within certain limits set by police, cases (calls to a home) were randomly assigned to three different methods of handling: arrest, mandatory mediation, or immediate temporary separation. The main object was to determine which of these regimens produced the lowest level of subsequent domestic violence in households. The integrity of the experiment was sustained by the randomization. That is, families involved in cases in each group were equivalent on account of the random assignment. Competing explanations common in earlier nonrandomized studies could be ruled out. Such competing explanations included differential police preferences for one or another way to handle the violence complaint. The experiment helped to inform a fifteen-year debate on handling such cases and is likely to be repeated in other cities to ensure the generalizability of the findings. ADMINISTRATIVE LAW: TELECONFERENCING. Most administrative appeals hearings for unemployment insurance and welfare are conducted faceto-face. However, travel distance, workload of hearing officers, schedules, and so on, are often impediments. Distance is a special problem in the Western states. One could argue that telephone hearings are a viable alternative in that they can result in more timely appeals and reduced costs. Experiments by Corsi and Hurley (1979a,b) and Corsi (1983) were designed to compare the efficacy of hearings conducted over the telephone compared with those conducted in person. The experiment was conducted statewide in New Mexico during the late 1970s.

346

Appendix B

Individuals were randomly assigned to telephone versus in-person hearings. The legality of the random assignment strategy was examined in detail, based on court decisions (e.g., on opportunity to cross-examine, on due process), state statutes on conduct of hearings, and so on. The Legal Aid Society, the relevant attorneys, and others were involved in the examination and the actual design of the study. Special provision was made for exceptional claimant cases. The experiment is remarkable in being the first conscientious attempt to understand the effectiveness of telephone versus in-person hearings. It is remarkable for its illustration of the feasibility of such a test in a complex, potentially controversial, and certainly sensitive environment. It is pertinent to compliance experiments in each respect. OTHER RANDOMIZED EXPERIMENTS. The bibliography to this appendix lists experiments conducted in other relatively sensitive areas. Rossi, Berk, and Lenihan (1980), for instance, report on statewide experiments in Texas and Georgia to determine whether post-prison financial support reduces recidivism. Randomized field tests in the courts are described by Goldman (1977, 1985) for pretrial hearings and Lind (1985) for assessing court-annexed arbitration, mediation in appellate courts, and others. Juvenile programs including restitution have been examined in randomized experiments by, among others, Lipsey, Cordray, and Berger (1981) and Severy and Whitaker (1982). Earlier work is listed in a bibliography by Boruch, McSweeny, and Soderstrom (1978). More recent reports can be found routinely in journals such as Evaluation Review, New Directions for Program Evaluation, and Program Planning and Evaluation.

Feasibility: Issues and Criteria The scientific justification for randomized assignment, of course, lies in generating a less equivocal, unbiased estimate of the relative effects of a compliance program, an estimate that is coupled to a formal statement about one's certainty of the results. Such a justification is important in the sense of determining when an experiment is appropriate. When framed persuasively, it may also enhance feasibility. Four broad categories of feasibility issues are considered in this section, including the experiment's justification: • Legal and ethical issues; • Scientific issues;

Experimental Designs

347

• Political and institutional issues; and • Managerial issues. These categories comport with those in Riecken et al. (1974). L E G A L AND E T H I C A L ISSUES

RANDOMIZATION. T w o themes are important. First, there is a need to distinguish between legal problems engendered by randomization and the difficulties of judging moral correctness of activities. Little can be gained by combining the two, and at worst, it may confuse matters. Randomizing at the individual level, for instance, may be legal. But the act may be unethical in the sense that it engenders great discomfort for certain respondents or in the sense that responses cannot be protected from uses that harm the respondent. For illustrations and analysis by an able legal scholar, see B r e g e r (1983).

The second theme is that there are multiple approaches to resolving legal and ethical problems in randomized experiments. Each has some costs and benefits. The approaches considered below include legal and technical approaches to resolving problems in two areas: randomization and privacy. The last part of this section concerns institutional review boards, a useful administrative device for ensuring that community values are represented in any assessment of the ethical propriety of a social experiment. See Boruch and Cecil (1983) for more on the multiple solutions theme. LAW AND RANDOMIZATION. The IRS's administrative authority for conducting experiments, Congress's role in specifying statutes that encourage tax experiments, and the courts' interpretation of each are especially important. The following brief review is based on Breger (1983) and on the Federal Judicial Center (1983); these papers should be consulted for legal citations. In general, the administrative authority for conducting experimental tests of federal projects stems from the enabling statutes of the various federal agencies. The statutes may be specific, as in the case of law that directs the U.S. Department of Housing and Urban Development to undertake housing allowance experiments. Or the law may be general, as in the case of statutory provisions that an agency head may waive compliance with other statutes for purposes of experimentation or demonstration projects, as, for example, the waiver authority of the secretary of the U.S. Department of Health and Human Services that permits social experimentation, within limits. Administrative authority to test alternative ways to increase compliance

348

Appendix B

has, of course, been exercised by the Internal Revenue Service. The Long Range Tax Forms Simplification Study (IRS, 1983) included two major alternative tax form packages. The development effort also considered a simplification, a third form suggested by discussion with the IRS commissioner, that led to the 1040EZ. A noteworthy statutory provision, Section 155 of Public Law 95-600, provided explicit authority to the IRS for contracting the job of redesigning forms. Federal authority for randomized field tests of social intervention programs has been challenged in the courts on at least two occasions. The courts have interpreted the authority broadly, rejecting the challenges to experiments in Aguayo v. Richardson and California Welfare Rights Organization (CWRO) v. Richardson. It is worth remembering more generally that the legal history of experimentation—experimentation being defined broadly—is compatible with a variety of court decisions. In the Truax v. Corrigan decision of 1914, for example, Justice Holmes rendered a dissenting opinion that "there is nothing I more deprecate than the use of the Fourteenth Amendment beyond the absolute compulsion of its words to prevent the making of social experiments . . . in the insulated chambers of the several states, even though the experiments may seem futile or even noxious to me and to those whose judgment I most respect." A similar spirit is reflected in comments by Justice Brandeis in New York State Ice v. Liebmann. The use of a safety-valve category in randomized tests has been a device to ensure that unfairly harsh or severe burdens are not imposed on certain individuals (or institutions) as a consequence of the experiment. That is, providing for special exceptions to randomization can reconcile institutional responsibility to take action on the basis of special needs or conditions of the individual in law, medicine, and elsewhere with the need for fair assessments of innovative practices. The categories to be excepted should be specified beforehand; e.g., the Sherman-Berk police experiments recognized the need to provide for police discretion in handling particularly violent domestic battles. Of course, providing ambiguous or large exception categories can undermine experiments; e.g., judges may overuse exceptions in a court experiment (Conner, 1977). T E C H N I C A L DESIGN APPROACHES AND RANDOMIZATION. There are a variety of strategies for tailoring experimental designs so as to avoid legal and ethical problems. They are as relevant to compliance experiments as they are to research in other areas. One obvious tactic is to ensure that the number of individuals (or other

Experimental Designs

349

entities) exposed to potentially burdensome treatments is kept to a minimum by choosing the appropriate sample size with statistical advice and by coordinating efforts to avoid unnecessary redundancy in risk-laden studies. The technology for minimizing the number of questions that must be asked and minimizing sample size under a variety of constraints is adaptable from sample design technology developed to reduce costs without appreciably reducing precision in research. Second, testing components of programs rather than full programs, for example, is often warranted on management grounds, since full-blown tests are expensive and may not be worth the effort, as well as on ethical or legal grounds. The Perng (1985) experiments, for example, focus on letters as a device for encouraging compliance, rather than the broad range of communication devices available to the IRS. Similarly, variations in program intensity may be compared to skirt the ethical problem of testing a program against a no-program control condition. Third, one may alter the units of randomization and analysis, randomizing institutions when that process is legal and ethical and randomization of individuals is not. This approach puts stress on direcdy estimating a new program's impact on institutions. Experiments that use as the unit IRS offices, or special groups within each office, will at times be more feasible than experiments that randomize individuals. Graduated introduction or withdrawal of benefits whose effect is not clear can be tailored to designs that capitalize on the stagewise character of the process. These are discussed below in the special context of compliance work. PRIVACY. Private information is defined here as information that, if disclosed, would bring harm to the individual. The term confidential refers to the state of the information, notably the fact that it is not disclosed or should not be disclosed. These definitions, adopted from suggestions by Albert J. Reiss, Jr., improve on earlier ones (Boruch and Cecil, 1979). The approaches described below ensure that private information once collected will remain confidential. Some methods also protect privacy in the sense that not even the researcher knows the true status of the research participant. The functional distinction between research information and administrative information is often critical to law, regulations, and development of privacy protection methods. Administrative information is defined as information used in making judgments and decisions about specific identifiable individuals. Research information uses individuals' identification only as a

350 Appendix B tracking device. It is not material to the conclusions, usually statistical, drawn about the group to which identifiable individuals happen to belong. This distinction can become blurred for compliance research. Indeed, the conduct of legitimate social research by an enforcement agency such as the IRS (or state tax bureau) itself presents some interesting ethical problems. They are discussed briefly below. The first major theme of the discussion is that there are multiple solutions to protecting privacy or confidentiality of individual responses in compliance research. The classes reviewed here are procedural, statistical, and statutory solutions. The second theme is that the risk of forced or accidental disclosure of sensitive information is very low. Fewer than a dozen subpoenas have ever been issued; few cases of accidental disclosure have ever been discovered. This implies that a sensible balance must be found between the cost of privacy protection and (usually weak) threats to privacy. The material is adapted from Boruch and Cecil (1979) and elsewhere. PROCEDURAL APPROACHES TO PROTECTING PRIVACY.

T h e s e are n o n -

technical approaches to assuring confidentiality or privacy. In mail surveys, for example, anonymous responses and alias responses (when the study is longitudinal) are not uncommon approaches. Inquiry that is indirect may be made through brokerage agencies, an archive or the Census Bureau, for example, to obtain data without necessarily obtaining identification, for both cross-sectional and longitudinal studies. Insulated data bank approaches may be used to link records from different archives without breaching privacy rules governing each archive. The strategies are imperfect to the extent that deductive disclosure is possible and the procedures are burdensome and time-consuming. Deductive disclosure seems critical to compliance research and the work by Schueren (1985) among others should be explored. At times, the more important cost is degraded quality of research, since it may be difficult or impossible to assay the validity of the sampling and the quality of the response. STATISTICAL APPROACHES TO PROTECTING PRIVACY.

O v e r the past ten

years, a variety of clever statistical devices have been developed to obtain information from identifiable responses in direct surveys without degrading either privacy or quality of research. One such approach has in fact been tested by the IRS in field studies of opinion about tax cheating (Aitken and Bonneville, 1980). A simple variation involves presenting two questions to a respondent,

Experimental Designs

35i

one sensitive (e.g., Did you avoid paying legitimate taxes last year?) and an innocuous one (e.g., Did you give to a charity last year?) that is independent of the first. The respondent is asked to choose one or the other based on the roll of a die and to answer it without disclosing which question is being answered. The respondents might be instructed to address the first question if a i appears on the thrown die, for instance, and to address the second question if a 2 or 3, etc., appears. The interviewer receives only yes or no responses and does not know which question was answered. Given two large samples in which the odds of answering each question differ, and the odds of answering each question are known, the statistician can legitimately estimate the fraction who avoided paying legitimate taxes. Yet privacy is protected in the sense that no unambiguous sensitive information is disclosed. Such devices amount to having the respondent inject his or her response with random error so that the interviewer cannot link a response with a specific question or state of the individual. Tested in the United States and abroad, the methods have considerable promise and appear to work well often but not always. The methods have failed, for instance, in that at times they produce negative estimates of the incidence of the trait of interest. Part of the problem appears to be that some respondents view a "no" as the safest possible response, that is, they do not conform to instructions despite the logical guarantee that sensitive information remains private. The failure to conform and the kinds of formats and questions that increase willingness to abide by instructions need a good deal more research. Even when the methods work well, the direct costs of their use increase in complexity of analysis and in required sample size. They are useless with small samples and with clinical inquiries. The implications of the new work in this area for compliance research need to be examined. STATUTORY APPROACHES TO PRIVACY PROTECTION. Statutory approaches

refer to laws that state that identifiable information collected for research purposes cannot be used for judicial, administrative, legislative, or other purposes against the individual. The level of protection available varies by statute. In some cases, identification and information are protected; in others, only identifiers are protected. Statutes have been created to protect respondents in federally supported research on criminal justice, mental health, and drug and alcohol abuse. A few court cases have helped to clarify the limits of their protection. The laws apply to data on individuals, not institutions, and this may limit their usefulness in compliance research, for example, on small businesses.

352

Appendix B

The statutes most pertinent to compliance research include: • Public Health Services Act, P.L. 93-282, Par. 303; • Crime Control Act of 1973, P.L. 93-83, Par. 524(a); and • Drug Abuse Office and Treatment Act, P.L. 92-255, Par. 408, and there are others. To illustrate, Section 524(a) of the Crime Control Act says that persons receiving federal research grants or contracts under the Act shall not reveal research information identifiable to any person for any purpose other than that for which it was collected. Copies of such information shall be immune from legal process, and shall not, without respondent consent, be admitted as evidence or other purpose in judicial or administrative proceedings. Evaluations of the statutes are given in Nelson and Hedrick (1983). There appears to be no focused investigation of the extent to which such statutes are applicable to tax administration research. However, the statutes are broad enough and sufficiently well tested in the courts to warrant attention in this context. In particular, they may help to elicit more candid reporting in research on nonreporting and dishonest reporting by reducing the respondents' concerns about the research and to encourage the cautious researcher to enter tax compliance research by ensuring that he or she can adhere to promises of confidentiality. An institutional review board (IRB) is a group formally convened to review the ethical propriety of research on human subjects carried out at an institution. In biomedical and behavioral research, for example, IRBs are required by federal regulation (45 C F R 46) to oversee the rights and welfare of individuals involved in such research, risks and potential benefits of the work, and appropriateness of methods used. The regulations require diversity in committee composition; e.g., no committee or quorum of a committee can consist of employees of the institution or of members of a single professional organization. INSTITUTIONAL REVIEW BOARDS: GENERAL PROTECTION.

The regulations generally recognize lower risks typically accruing to social research in contrast to medical research. Expedited review processes are possible for the former and for innocuous varieties of the latter (e.g. cuticle samples). The performance of IRBs has been assessed intensively since 1974 by a variety of organizations. Regulations have been changed during that period by the Department of Health and Human Services and the Food and Drug Administration among others to strengthen protection

Experimental Designs

353

and reduce unnecessary constraints on the research (see Levine, 1981, for example). The legitimacy of IRBs has been reinforced partly by the federal courts. In Crane v. Mathews, for instance, the court ruled that an experiment involving medical insurance copayment plans did indeed involve human subjects at risk and required that the research be assessed by an IRB. In doing so, the court ruled against a 1976 position taken by the Department of Health, Education, and Welfare (Breger, 1983). More recent rulings have further clarified the IRB role. The IRB approach to ensuring propriety of experiments is relevant to tax compliance research for several reasons. First, the IRBs are designed to cover the propriety of research on human subjects. Tax compliance research, including experimental tests of compliance programs, does involve research on human subjects. Analogous research, such as the negative income tax experiments, insurance copayment experiments, and administrative hearings experiments have been routinely assessed by IRBs in the interest of protecting subjects' rights. Second, IRBs are empowered to review a great deal of work that is at least as sensitive as tax compliance research. This includes medical experimentation, of course, and social experimentation on topics such as medical insurance copayments, prison sentences and police methods, white collar crime and deception in surveys, divorce and reconciliation, and so on. Third, IRB performance has since the 1970s been clarified well by federal agencies, examined often and found appropriate by independent groups, and improved with respect to quality of review. Tax compliance research can capitalize on this administrative experience. Fourth, the IRB reviews of social experiments of the sort the IRS might run, in contrast to medical experiments, have been made much more efficient over the past few years. These improvements have taken the form of exemptions, expedited review procedures, and waivers of informed consent requirements for certain classes of studies. It is likely, for example, that waivers would be relevant for relatively innocuous administrative experiments involving, say, different letters sent to delinquent taxpayers. The experiments that put subjects at plausible risk but with plausible benefits to society are most likely to be legitimate subjects for IRB attention. Information about institutional review boards appears frequently in publications such as IRB, issued by the Hastings Center's Institute of Society, Ethics and the Life Sciences (360 Broadway, Hastings-on-

354

Appendix B

Hudson, NY 10706). Recent developments in the federal sector provide an opportunity for the I R S to explore the use o f IRBs and ethics regulations in its own work. In particular, the Interagency Human Subjects Coordinating Committee consists of about seventeen federal agencies dedicated to developing a model policy for human subjects research. The committee is chaired by Charles McCarthy o f the National Institute of Health's Office for Protection from Research Risks. Joan Porter, Special Assistant to McCarthy, was kind enough to provide information about the committee and an updated copy of the regulations regarding human subjects research. The committee's draft model policy is similar in many respects to the regulations, but it has not been formally approved by all the agencies contributing to it. Definitions o f research, human subjects, and other concepts in both model policy and current regulations are as pertinent to the IRS's contracted research as to other federal agencies' work. It is in the interest o f agencies such as the I R S to learn about the interagency committee. It is clearly in IRS's interest to consider adoption o f existing regulations or model policy in order to ensure the ethical propriety o f its own work regardless o f whether the work involves randomized field tests. It is also in the IRS's interest to learn about the committee insofar as exemptions justifiably remove some research from I R B review.

SCIENTIFIC ISSUES

A variety o f scholarly review papers in various disciplines display no essential disagreement about the scientific merit o f randomized tests in principle. This includes Farrington (1983) on criminal justice, Rossi and Freeman (1981) on health services research, and Ferber and Hirsch (1982) and others on employment and training experiments. Differences do appear among the disciplines in scholarly debate about the extent to which alternative methods produce equally persuasive evidence. In particular: Do other statistical methods produce as accurate an estimate o f project effect as a randomized experiment? That is, what do we know from empirical comparisons? The question is as relevant to tax compliance research as it is to other areas. EMPIRICAL COMPARISONS OF RANDOMIZED AND NONRANDOMIZED TESTS.

The idea o f comparing outcomes o f a randomized experiment to results achieved in nonrandomized trials is not new. T o clarify arguments with Gosset in the 1930s, for example, Fisher appears to have tried such a comparison in experiments on wheat (Box, 1978:269) and in reanalyzing

Experimental Designs

355

Darwin's data on stock fertilization. In the first case, results of the experiment differ from those of the quasi-experiment. In the second, they do not. Similar comparisons have been undertaken by medical researchers. Randomized trials on the Salk vaccine, for example, gave estimates of the vaccine's effect that differed from estimates based on nonrandomized quasiexperiments (Meier, 1972). The debate over when randomized tests are appropriate in medicine is not new, in the U.S. at least (Freund, 1968; Gray, 1975). What remains to be done in the medical arena includes increasing capacity and willingness to do experiments, if we may judge from the Institute of Medicine's (1985) report on postmarketing surveillance for drugs and other technologies. The tradition of evaluating employment and training programs has been based on econometric models and survey data. That is, in the absence of a randomized trial, one posits an explicit statistical model on which to base estimates of program effect. It is relevant to compliance research in that similar models and estimation techniques are often used to analyze survey data on compliance. That conventional econometric models are often wrong seems clear from recent work by Fraker and Maynard (1984; but see Heckman et al., 1987). In particular, a broad array of such models produced results that differed widely from randomized field tests of programs on similar populations for youth. The results did not, however, differ appreciably for samples of welfare-supported women with children. The important lesson here, as in the Salk trials and in other cases (Boruch, 1976), is that one cannot be sure beforehand that a nonrandomized design will produce the same results as a randomized set-up. The same lesson can be drawn from studies of groups of program evaluations. Gordon and Morse (1975), for example, found remarkably different estimates of social program effects when the evaluation design was taken into account. Glass and Smith (see Light and Pillemer, 1984) found differences in tests of the effect of class size on students' learning, when the quality of experimental design was taken into account. In a fascinating series of papers on medical experiments and quasi-experiments, Chalmers and his colleagues find analogous differences. Little recent evidence of a similar sort, however, appears to have been generated in criminal justice, law enforcement, or compliance research. The randomization procedure avoids the need for elaborate statistical models whose assumptions, though explicit, are often untestable. The batde between modelers and randomizers began in 1935 between Fisher

356

Appendix B

and Neyman (Box, 1978:265). It continues insofar as modelers are willing to trust their model and cannot invent (or imagine) experiments that help to test that trust. It continues insofar as the experimentalists distrust the models, plump for experiments, and do not have the resources to conduct the experiments that test the models. The implications of all this for compliance research and tax administration research more generally are that: 1. One cannot predict well that an experiment will produce the same result as a quasi-experiment; 2. Empirical comparisons are desirable to understand when the different methods are likely to converge; and 3. Calibration experiments are likely to be a useful device for gauging the appropriateness of nonrandomized trials. "Calibration experiments" here mean set-ups in which economical methods for nonrandomized tests are run side by side in various contexts with randomized tests in order to estimate the biases engendered by the former and to use these estimates in other similar contexts that do not permit randomization. POLITICAL AND INSTITUTIONAL ISSUES

A policy experiment is, as Riecken and others (1974) observe, a political act. It often demands recognition of political and institutional realities, including the need to negotiate with a variety of stakeholders. The Internal Revenue Service is positioned well to avoid political and institutional problems insofar as it confines attention to testing important low-visibility administrative innovations. The Accounts Receivable Experiment and the randomized response experiments described earlier illustrate the genre. Tests of more visible innovations are likely to be a good deal more controversial, and as important. Learning to cope with predictable problems seems sensible. Consider, for example, the use of telephone hearings instead of inperson hearings on administrative appeals. The early discussions leading to Corsi's (1983) administrative law experiment concerned whether due process in welfare appeals cases was indeed met through telephone hearings instead of the more conventional in-person hearings, whether administrative agencies had authority to use the innovative telephone approach, and whether random assignment was really necessary to compare telephone with in-person hearings. The negotiations involved a half-dozen government agencies, assorted public administrators, and lawyers; discussions gradually clarified concerns of these stakeholders. Sherman and Berk (1984) focused on the law enforcement decision

Experimental Designs

357

chain in the Minneapolis experiment. The sponsoring agency's concerns about ethical issues were handled partly through discussion of the sparse evidence concerning police effectiveness in handling domestic violence. The objections of interest groups, such as the American Civil Liberties Union and women's rights groups, had to be met with evidence and argument, too. Mayoral support in two administrations had to be developed and sustained. Meetings and retreats with beat cops and their supervisors appear to have been frequent and effective in handling their concerns. The Work/Welfare Demonstration (Gueron, 1985) also demanded considerable negotiation and planning, partly because eight state governments were involved. The Manpower Development Research Corporation's negotiative work appears to have been productive by capitalizing on local expertise to nail down details, introduce good ideas, and ensure realism. PUBLIC ATTITUDES ABOUT RANDOMIZATION.

Certain attitudes can be

regarded as indicators of social ethics. For instance, do individuals involved in experiments find randomized assignment objectionable or demeaning? What do we know about attitudes, preferences, and opinions or about the accuracy of information on which the opinions are based? In fact, we know little—but evidence is accumulating. Laboratory studies on attitudes, for instance, are just beginning. In Australia, J. M. Innes has executed small studies to understand how individuals view random assignment to new family therapy programs when various justifications for such assignment are stressed to the eligible families: scientific need for evidence on the program's effect, the equity of randomization when resources are scarce, and the possible negative effects of innovative treatment. The results suggest that the appeals to scientific or equity arguments do not appreciably affect a favorable attitude toward randomization but that the possibility of negative program effects on participants does. Studies by Hillis and Wortman (1976) in the United States on medical experiments suggest that randomization is indeed viewed in a more favorable light when scientific merit of the experiment is emphasized and that it is viewed in a less favorable light than are alternatives when resources are scarce. Subsequent work by Boruch, Dennis, and Greer (1988) and Boruch and Cecil (1983) laid out problems and alternative solutions in experiments that engender ethical or professional issues. EXPERIMENTS IN EXPANSIONIST AND REDUCTIONIST PUBLIC POLICIES.

The 1960s and 1970s were expansionist in the sense that the U.S. government provided increasing services to the disadvantaged. The Great Society

358

Appendix B

ideas were often innovative. When there was opportunity to innovate, there was also opportunity to experiment formally, to assess the consequences of the innovation. And in fact, major economic experiments were undertaken during this period. This interest in assessment is a fine scientific rationale for experiments. It may also serve the scholar-bureaucrat well. The rationale does not appear to be sufficient early in an expansionist period, however. Rather, the demand for evidence from the conservative camp appears to have led to demands for better evidence. For the cautious legislator, evidence from field experiments appears to have played a role in deciding whether and how to support or oppose legislation. During a conservative regime there is more likely to be interest in learning more about the effectiveness of new variations on existing programs and new approaches to cost reductions, law enforcement, and compliance-related programs. Some of the pressure to present evidence is likely then to come from the opposition party. Here, too, randomized experiments are a natural vehicle for obtaining relevant evidence. M A N A G E R I A L ISSUES

"Capacity" here means access to competent staff, resources such as money and time, the control that is needed and can be exercised to experiment, and the tolerability or tractability of constraints on the randomization process. Staff. N o good experiments are mounted without an able cadre that has some assurance of sustained support. In the United States, the resources expended on program evaluation generally have led to private, for-profit, and not-for-profit organizations' developing the requisite skills. These organizations engage in a competitive bidding process to provide the human resources necessary to plan and execute the field experiment. Some are specialized. The Manpower Development Research Corporation (Gueron, 1985) and Mathematica Policy Research (Hollister, Kemper, and Maynard, 1984), for example, dedicate their attention to experimental tests of human resource programs. The Police Foundation and the Crime Control Institute in Washington, D.C., have undertaken tests of police patrol strategies and alternative ways of handling domestic violence (Sherman and Berk, 1984). Organizations that cast a wider net in their applied social research, such as Abt Associates, have done fine experimental tests of alternative ways to reduce the costs of day care programs for preschool children and to find

Experimental Designs

359

effective methods of nutrition education (St. Pierre and Rezmovic, 1982; St. Pierre et al., 1982). Rand's health insurance experiments are well known (Brook et al., 1984). S R I has had considerable experience in the negative income tax arena (Robins et al., 1980; Robins and West, 1981). These are supplemented often and displaced at times by university-based groups. CONTROL AND CONSTRAINTS. An important lesson of Conner's (1977) study of twelve field experiments is that controlling the randomization process counts heavily in a successful social experiment. Control is no less important in medical randomized trials, of course. Friedman, Furberg, and DeMets (1985), for example, were careful to encourage blind randomization in drug effectiveness experiments that might otherwise be easily subverted by medical specialists involved in the trials. The lack of control over randomization accounts at least partly for failures of T V experiments in El Salvador (Hornick et al., 1973), the Roos, Roos, and McKinley (1977) trials in health services, and the Bickman (1985) tests of nutrition programs for the elderly and others. The mechanism for exercising control varies considerably. In the I R S taxpayer compliance experiments described by Perng (1985), centralized control over randomization is a natural part of the administrative process. More loosely coupled social systems are more typical, however. The need for control in looser contexts has led to the use of centralized random assignment (by telephone) in the work/welfare tests that Gueron (1985) describes. That is, welfare program staff provided some client information to M D R C , and assignment to alternative regimens was made from random number tables constructed to generate blocks and groups. Telephonebased blind assignment was also used by Goldman (1977) in experiments on how to reduce time in court with pretrial hearings. The Police Foundation's approach has been less direct in that police officers were responsible for assigning on the basis of a "randomization pad." But surveillance was possible to ensure adherence to the rule (Sherman and Berk, 1984). Exercising control also implies limits on control. A physician's relinquishing direct influence over the treatment of a patient, for instance, presents a major ethical problem. The same problem, and tension over diminished or shared professional power, is evident in judicial experiments (Goldman, 1977), police experiments (Sherman and Berk, 1984), educational experiments, and others. Exclusionary rules are crucial to discretion for the courts (for example, the pretrial hearings experiments), to program staff (for example, the Minneapolis Domestic Violence Experiments) or to ensure compliance with statute (Breger, 1983) or regulation. The main

360

Appendix B

lesson of all this is that exclusionary rules must be specified beforehand or as early as possible in the experiment. RESOURCES: TIME AND MONEY.

N o general information has been

compiled on how much experiments have cost. No journal articles or books have reported detailed analysis of their costs. This prevents benefit-cost analysis of individual experiments and comparing relative costs and benefits across disciplines (e.g., Mosteller and Weinstein, 1985). All this is despite the Social Science Research Council's efforts, among others, to encourage such reports (Boruch, 1976) and complaints of economists, such as Nobel laureate Theodore Schultz (1982) that economists do no benefit-cost analyses of their own work. There are a few exceptions, notably the Abt Associates' analysis of day care experiments. The experiment itself concerned whether relaxing federal regulations on the credentials of the day care staff would have negative effects on children. The results were used to alter regulations. The use resulted in the experiment's paying for itself in less than five years.

Special Design Issues There are a variety of issues involved in experimental and quasi-experimental design that are likely to be important in compliance work. Some of the issues have already been discussed. A few deserve more attention and they are discussed here. UNITS OF RANDOMIZATION AND ANALYSIS

The units of randomization can at times be chosen so as to facilitate field tests. To judge from the educational research, for example, students within classrooms cannot be randomly assigned to one of two or more regimens. In contrast, the unit that can be feasibly randomized may consist of a group of individuals—a neighborhood, a family, a classroom, entire schools, hospitals, or other institutions. In fact, the tactic of randomizing aggregates has been exploited to execute experiments in Colombia, where small, randomly assigned geographic sectors have been used to estimate the effects of preschool and nutrition programs, and in Nicaragua, where classrooms were assigned to alternative regimens to estimate the effects of radio-based mathematics education. Bickman's (1985) tests of children's health education and school improvement incentives involved random assignment of schools; his tests

Experimental Designs

361

of nutrition education for the elderly involved assignment of sites to alternative regimens. That the choice of unit can indeed be difficult is clear from Sherman and Berk's (1984) consideration of alternative units for randomization in the Minneapolis Domestic Violence Experiment: days, precincts, cases within individual officers' workloads, officer within treatment versus officer across treatment. Their resolution was randomizing case handling method within officer. It appears to have worked very well, without appreciable loss of experimenter control. As Lind (1985) points out, there are similarly difficult choices in court experiments, too, notably in deciding whether individual cases or related cases ought to be taken as the unit. The choice of unit seems relevant to compliance research. That is, individuals within local office, local office, taxpayers, small groups of taxpayers within strata, and so on, may lend themselves differentially to experimentation. In fact, the compliance experiments about which we know use only the individual taxpayer as the unit of randomization and analysis. This is the right unit given the objectives of the studies reviewed earlier on delinquent accounts, simplification, and so on. Individuals do not present themselves for random assignment all at once, of course. The little bunches that appear may themselves be randomized. A variation on this theme is time-bound quota sampling used by Hillsman-Baker and Rodriguez (1979) in their court experiments. Singlecase randomized assignment was unacceptable to lawyers involved in the tests, but there were no objections to quota sampling when there were more cases than could be assigned to treatments. The investigators randomized time segments that determined how large a quota could be expected. In consequence, the first intakes in a segment were assigned to treatment until the quota was filled; the remaining intakes were assigned to control. Segments of different lengths were randomly arranged over person-work-hours to reduce periodicity. Periodicity and "suitable" number for randomization within subgroup can be manipulated within reasonable limits. And they influence feasibility and integrity of randomization. For instance, Collins and Elkin (1985) take seriously the question whether to randomize within groups of four individuals (to each of four treatments) or to randomize within groups of sixteen. Either approach was feasible given the client flow; either could achieve numerical balance across treatments. The sixteen-member subgroup was chosen eventually to make subversion of the randomization very difficult. That is, it is easier to anticipate where particular patients will wind up when

362

Appendix B

groups of four are randomized in sequence than it is in the case of sixteen. The tactic of randomizing within what is, in effect, a time block is not unusual. See for instance, Hahn (1984) on acoustic experiments with airport birds. The stage at which randomization occurs is also a research design parameter for able managers of field tests. So, for example, randomized assignment of all individuals eligible for special services for the chronically ill may not always be sensible prior to eliciting their consent to participate in the trials. Random assignment after eligibility screening and after consent is elicited generally reduces the sample and the generalizability of results. Despite this, the screening is sensible if the object is to get accurate results at slight reductions in generalizability. So, for example, Bickman, among others stresses the need to delay random assignment until the last possible stage of an endeavor, mainly to avoid complicated attrition problems.

R E A C T I V I T Y A N D V A L I D I T Y OF T H E E X P E R I M E N T S

Reactivity here refers to an experimental subject's awareness of the experiment and the effect of that awareness on his or her behavior. So, for example, an individual who knows the pill is experimental may be inclined to feel positively or negatively about it, regardless of the pill's physiological effect. The feeling undercuts efforts to assess actual medical effects. Because individuals react in a variety of unpredictable ways to their awareness, it is common practice in medicine to use placebos and doubleblind measurement. That is, the volunteer patient in a clinical trial is kept unaware of which of two potentially effective drugs is being administered. Those responsible for measurement are also blind to types of treatment. Awareness can plausibly be expected to affect validity of social experiments. But accommodating the problem is usually more difficult. Furthermore, the matter becomes very complicated if one considers awareness at each stage in the experiment from the random assignment through process and eventual measurement of response. Two broad questions are pertinent to this threat to validity of randomized tests: Is the threat plausible? Is its magnitude likely to be substantial? What can be done about reducing and estimating the threat's magnitude in the settings in which it is plausible? Is the threat plausible ? Clearly we can categorize compliance studies into those in which • Reactivity effects are implausible;

Experimental Designs

363

• Reactivity effects are unpredictable; and • Reactivity effects are plausible and likely to be great. Settings in which reactivity effects are least plausible include all those in which the subjects of research do not know they are involved in an experiment. As a practical matter this includes all experiments in which the treatments are a matter of small changes in administrative or bureaucratic processes that are not terribly visible to the public. Some examples include: • Tone, frequency, length, and style of telephone calls and reminder letters to delinquents or other subgroups, as in the Accounts Receivable Treatments Study; • Character of questions asked of a national probability sample in a poll about the IRS, as in the randomized response experiment; and • Single-shot, fast turnaround experiments in which word about the experiments cannot get out quickly, as in the Schwartz-Orleans experiment. The experiments that are most likely to involve reactive effects include those in which: (a) subjects are aware that their treatment is different from others, and (b) the difference is regarded as important enough to engender differences in behavior. For instance, a randomized test of the effect of amnesty programs on tax delinquents may engender distress or anger among control group members who have not been offered amnesty or who regard amnesty as unfair to honest taxpayers. That distress may lead such individuals to resist further, to seek remedies to perceived unfairness in the courts, or to take some other action. What can be done to reduce the threat ofreactivity? An obvious option is to choose to do only those experiments for which plausibility of occurrence and magnitude of reactivity are likely to be low. The next option is to capitalize on strategies that actively reduce the threat. In particular, one may: avoid publicization of the experiment till it is complete, refrain from informing subjects that they are in an experiment, design the experiment so that its conduct is fast and its visibility low ("stay low and keep moving"). The IRS Research Division has, in fact, avoided publicity in earlier tests so as to avoid distortion of results, notably in the Long Range Tax Forms Simplification Study (IRS, 1983). Regardless of which option is chosen, it is sensible to estimate the magnitude of the reactivity. In particular, it seems sensible to design a side experiment whenever possible to determine whether and how awareness

364

Appendix B

influences behavior. This requires that at least one subsample be unaware of the experiment and that awareness be measured. For instance, the IRS's (1983) Long Range Tax Forms Simplification Study included a control group that was unaware of any experiment. The return rate of taxpayers in each of the three experimental groups did not differ appreciably from this group's rate, suggesting that the experiment was not reactive in this particular respect. Similarly, Schwartz and Orleans included both a placebo group and an untreated control to understand the placebo effects of the experiment on taxpayers (no significant effect evidendy appeared). C O U P L I N G EXPERIMENTS TO SURVEYS AND QUASI-EXPERIMENTS

Randomized experiments can be coupled to longitudinal surveys and panel studies. The purposes of this "satellite" policy include: calibration of nonrandomized experiments, more generalizable randomized experiments, and better methods for estimating program effects. Longitudinal studies and surveys such as the TCMP based on welldesigned probability samples are clearly useful, for management and policy, in understanding how individuals (or institutions) change over time. For example, longitudinal studies avoid the logical traps that cross-sectional studies invite, such as overlooking cohort effects, in economic, psychological, and other research. Cross-sectional studies are useful for crude trend analyses. Both types of survey designs are often pressed, however, to produce evidence that they cannot support. Of special concern here is evidence about the impact of a compliance program on groups that a longitudinal or cross-sectional study happens to include. In employment and training work, for example, the Continuous Longitudinal Manpower Survey has been justified and supported primarily on grounds that we ought to understand what happens to the human resources pool. Its secondary justification is that it can help understand the effect of special programs—in youth employment, training, and so on. This secondary claim is unwarranted. Longitudinal and cross-sectional surveys are often not sufficient to permit us to estimate the effect of programs designed to, say, affect reported income of individuals who happen to be members of a survey sample, the crime rates of these people, and so on. That the claims made for longitudinal surveys with respect to evaluating effects of programs can be misleading is clear empirically and analytically. The most dramatic recent empirical evidence stems from Fraker and

Experimental Designs

365

Maynard's (1985) comparisons of program effects based on randomized experiments against effects based on nonrandomized data, notably the Continuous Longitudinal Manpower Survey and the Current Population Survey (but see Heckman, Hotz, and Dabos, 1987). Earlier evidence in different arenas stems from the Salk vaccine trials, health services research, and others (see Boruch, 1976, for a listing). Randomized experiments, by contrast, permit one to estimate the effects of projects with considerably more confidence. Indeed, reports of the National Research Council's Committee on Youth Employment Programs (Betsey et al., 1985), the Institute of Medicine's (1985) report on testing medical devices, and others are emphatic on this account. A major shortcoming of experiments, one not shared by the large-scale longitudinal studies, is their limited generalizability. That is, a set of experiments might be feasible in only a half-dozen sites, and those sites may not necessarily reflect national characteristics. The implication is that we ought to invent and try out research policy that couples the benefits of longitudinal studies, i.e., generalizability, with those of experiments, i.e., unbiased estimates of program effect. See Farrington, Ohlin, and Wilson (1986) for a recent treatment of the idea. The policy recommended here is akin to science policy on satellite use. The satellite, like the longitudinal or continuous cross-sectional survey, requires enormous resources to emplace and maintain. But scientists who design special-purpose work can obtain access to part of the satellite to sustain their investigation. Just as a physicist may use the satellite as a vehicle for limited, temporary investigation, the policy recommended here would allow the social science researcher to use the longitudinal survey infrastructure as a resource and as a vehicle for conducting prospective studies. This strategy engenders some obvious difficulties. For instance, learning how and when existing survey samples can be augmented for experiments or how and when subsamples of the main sample can be drawn off for experiments is not clear. Nor is it clear how to design the side experiments so as to avoid invidious demands on or damage to the main study. How real these difficulties are needs to be understood if any progress is to be made in this area. The best way to find out presumably is by trying the strategy out. The policy element gets beyond simple scientific traditions of data sharing (Fienberg, Martin, and Straf, 1985). It is considerably more debatable and more important in principle. Access is likely to be feasible, for example, for only a few projects, perhaps only one every year or two, on

366

Appendix B

account of the sheer difficulty of coupling studies to an already complex longitudinal enterprise. It deserves some thinking in the context of the TCMP.

ITERATIVE EXPERIMENTS: SEQUENCE AND THE LABORATORY-FIELD INTERFACE

A single-shot study, randomized experiment or otherwise, rarely produces results that are definitive or durable. It fails to be definitive often for reasons of imperfect execution and, more important here, because if done well it will raise new questions as well as answer older ones. In the social and behavioral arena, its results will be durable to the extent that laws and society change. A compliance program emphasizing threats, for example, may be cyclically effective as taxpayers grow less bold and more bold over time. Degradation of project effects occurs in long term tests of youth employment programs (Betsey et al., 1985) and elsewhere. One would expect some cycling and some degradation of the effectiveness in compliance work. The point, of course, is that a single-shot study may be useful for a problem at hand, but a sequence of experiments is more likely to yield results that have some generalizability in time. And, with the right designs, they will have generalizability with respect to populations of interest to the revenue services. Coupling of experiments and panel studies, as described earlier, is one way to ensure a more or less orderly sequence. Other approaches are possible. The interface between laboratory test and field experiments constitutes one aspect of iterative research. The idea is to cross-walk between laboratory and field periodically. The purpose is to capitalize on the inexpensive, controlled conditions of a laboratory at times and to exploit the field to verify laboratory work. Such cross-walking is not uncommon in research on jury decision making. Indeed, the IRS, in collaboration with private contractors, has used the strategy to understand how to simplify tax forms and whether laboratorybased results, on error rates and taxpayer burden, for example, hold up in large-scale field tests. In short, it doesn't take much wit to understand that sequence and a laboratory-field interface are sensible features of any experimentation policy. It takes considerably more, however, to determine how, how much, and when these features can be exploited.

Experimental Designs

367

BLOCKING VARIABLES A N D SPECIAL TARGET GROUPS

Certain groups of individuals may deserve special attention in experiments for two reasons. First, such groups may be especially interesting, because of their noncompliance rate, say, and so may be labeled high priority for attention. Second, and more important here, one group may react to a treatment regimen in ways that differ appreciably from another group's reaction. For instance, the Schwartz-Orleans results suggest that low-income taxpayers react in one way to appeals to moral conscience and high-income individuals react in another. Other, more recent studies also suggest that interactions between program type and person type are important and ought to be recognized in design work. Paid tax preparers, for instance, were very likely to choose conventional forms rather than experimental forms in the 1983 IRS Long Range Tax Forms Simplification Study. The individual taxpayers chose the (assigned but voluntary) experimental forms at a much higher rate. A similarly conservative choice strategy is also evident in subgroups of individual taxpayers. Lower-income, "married filing separately," and head of household filers were more likely than others to opt for the usual 1040A instead of the alternatives in this study. Schwartz and Orleans found tantalizing evidence for the idea that the attitudes of the highest socioeconomic class of taxpayers are notably affected by reminders about legal sanctions against tax compliance. The lower three classes, by contrast, were not affected appreciably. The lowest class category was affected most by appeals to moral conscience to judge by their responses to open-ended inquiries about reasons why taxpayers might report all income of a certain kind. Findings on employed versus self-employed taxpayers seem no less relevant to understanding which groups may lend themselves well to experimental tests of compliance programs. Small businesses are of special interest, of course, in that their noncompliance rate seems high; the self-employed are often small business owners. Moreover, the SchwartzOrleans work suggests that appeals to conscience rather than sanctions are likely to have a remarkable effect on compliance.

Quasi-Experimental Designs "Quasi-experimental design" here means a plan for collecting data that will yield estimates of program effect but does not include the randomized

j68

Appendix B

assignment of individuals (or entities) to one of several alternative regimens. These designs are labeled "quasi-experimental" in the educational, social, and psychological research literature and in some biostatistics work (Campbell and Stanley, 1963). Some of the designs are not different in principle from those employed in econometrics and biometrics. Only two such designs are considered here. They are demanding but are most likely to generate good estimates of the effectiveness of programs or projects. The two designs are regression-discontinuity approaches and time-series approaches. Other designs can produce defensible estimates of program effect, but the circumstances under which they do are more special and their assumptions are more demanding (see, e.g., Glass et al., 1975; Rossi and Freeman, 1981; Campbell and Boruch, 1975). D E S I G N S BASED ON PROGRAM ALLOCATED IN STRICT ACCORD WITH M E A S U R A B L E N E E D : REGRESSION-DISCONTINUITY

Randomized field tests are not always feasible. When people, groups, cities, or regions are assigned to receive services in strict accord with measurable need, however, a reasonable alternative design can be exploited. The design requires measurement of at least two points in time, on a sample of both service recipients and nonrecipients. The approach, developed by Donald T. Campbell (Riecken et al., 1974; Trochim, 1984) works in the following way. The need of regions or individuals is first measured; for example, compliance level is measured for each of (say) fifty sectors. The least compliant sectors are assigned to receive program services strictly on the basis of their measured level of compliance. For example, only the three with the lowest compliance rates receive program services or other treatment. Once the program has been emplaced and has had an opportunity to exercise notable influence, all sectors are measured again. The estimate of the program's effect hinges on the assumption that there is normally a strong, simple relation between compliance levels from one year to the next. To the extent that the program disrupts that relation, e.g., enhances the compliance rates in the areas subjected to special attention, the program can be said to have been effective. The estimate of program effect is then judged relative to a projection based on one prior measure and, as important, the assignment of sectors on the basis of measured need or merit. In the simplest case, the "strong simple relation" constitutes the null

Experimental Designs

369

condition—i.e., the standard against which program effects are judged. In the absence of a special program effect, the relation between before and after program observations is regular and, in this case, linear. If introduction of the program engenders a simple additive effect, then a disruption of the relation will appear. The effect of the program is reflected by the size of the discontinuity. This design is not yet commonly used. But is has substantial merit on technical grounds; the logic and statistical models reflecting the logic have been articulated well, by Trochim (1984), for instance. If the relationship of needs from one period to the next is strong and the program is assigned stricdy in accord with need, the design is at least as good, on technical grounds, as a randomized test. Its feasibility will be greater in some cases. See Trochim (1984) for illustrations and Rubin (1977) for a thoughtful analytic treatment. TIME-SERIES DESIGNS

In some program areas, good archival data are available to address particular interests. For example, accurate monthly records on the level of certain kinds of compliance may be available. The existence of such records makes a time-series approach to estimating program effects possible. In particular, individuals, organizations, or other entities in a target area are monitored at successive points in time before the program is introduced. The new program is put into place, and the individuals or entities are again monitored. The estimate of program effect is normally the difference between the level of compliance, for instance, following introduction of the program and the level evident prior to its introduction. The standard of comparison in this case is historical: a prediction, based on prior data, about what compliance would have been without the new program. Such historically based comparisons have merit to the extent that records are uniformly accurate and the time series is free from peculiar and poorly understood variation. To the extent that the quality of the records varies, or that factors other than the new program or project exercise unknown or incalculable influences on compliance, then estimates of the program's effect will be ambiguous at best and misleading at worst. Controlled experiments and time series are not inimical. They can be combined in an evaluation. The main point is that the randomized experiment sets up a contemporary empirical standard for judging the effect of a program. Effects are based on comparing a group of people who participated in Program A against an equivalent group who participated in

370

Appendix B

Program B. The time series involves a historical standard and assumption that is sometimes less tenable: that the behavior of the area or group under examination can be predicted confidently from earlier behavior alone. Because time-series analyses depend heavily on archival records whose quality is uniformly good and collected over a long period of time, the approach is adaptable only in particular areas and often not in new contexts. Good illustrations of time-series analyses of social data are not hard to identify: e.g., Glass et al. (1975), Nelson (1973), McCleary and Hay (1980), and Box and Tiao (1975). To determine the effectiveness of new, stringent laws governing drinking while driving in some parts of the United States, for example, time-series data on traffic accidents, arrests, and the like have been used. Variations in the quality of crime records and temporal variation in the definition of crime make it more difficult to use conventional administrative records for determining the effects of new laws on crime rates, but some attempts have been successful, e.g., estimating the effect of more severe sentences on carrying an unlicensed pistol in the state of Massachusetts. Analogous approaches to estimating the effects of new laws have been undertaken in Denmark. For example, the effect of relaxing pornography laws in Denmark was established in part by examining time-series data on sex-related crime rates. The Kutchinsky studies are especially conscientious because he recognizes that variation in legal definitions of crimes, cultural attitudes, and propensity to report crime is considerable, and he documents the extent to which these are plausible influences on the outcomes.

Some Crude Ideas About Field Experiments in Taxpayer Compliance EDUCATION AND EXPLANATION

Explaining what tax revenues are used for is held to be important by some as a device for increasing revenues. For instance, the Massachusetts Department of Revenue puts stock in the idea that voluntary compliance is enhanced by telling taxpayers that revenues are used for the Office of Children, the state police, homeless shelters, prison systems, environmental quality engineering, and other systems. How much effect this has is not clear, although some effect is plausible. The state of Massachusetts does not know how much effect explanations of this kind have, despite the importance attached to them. Experiments using the individual as the unit of

Experimental Designs

371

randomization seem feasible in view of precedents such as Schwartz and Orleans and Perng. Educating small businesses seems important from work in Massachusetts under Ira Jackson's leadership, work by the U.S. Congress's Joint Committee on Taxation, and others. The percentage of revenues voluntarily reported as a function of revenues that should have been reported is less than 50 percent according to the joint committee's report. So-called businessmen's kits and other educative consumer services directed at small businesses seem worth trying out. In particular, it is not clear how to best inform and remind new small businesses. Randomized tests using businesses as the unit may be warranted to test out an array of telephone, letter, and document-based advice. Similarly, accountants and firms that service small businesses might be regarded as a legitimate target population insofar as they provide poor or ambiguous advice to clients. These entities would then be the unit of randomization. M O N E Y AND REWARDS

The idea of creating explicit well-advertised rewards for taxpayer compliance has been taken seriously by the state of Massachusetts, among others. For instance, the state puts considerable stress on prompt refunds and advertises the fact. Rewards such as this are relative, of course; taxpayers who grow accustomed to fast refunds may in time be less affected by this incentive. Determining whether and when new kinds of monetary rewards have an effect on compliance seems important. One can conceive of a lottery, for instance, to ease the pain of being (randomly) audited. Small controlled tests of lottery-based reward in this and other contexts would help to establish whether the lottery-based rewards decreased indifference, hostility, or gratuitous criticism. (Deciding whether such attitude changes lead to more compliance is a separate matter for research.) ASSISTANCE AS R E W A R D

At least some individual taxpayers use the I R S to store cash in a sock. That is, they declare no deductions, have a maximum federal tax withheld, and usually wind up with a refund (without interest). This "enforced savings" strategy is of some benefit to the taxpayer, but it does reduce the amount of money available during the year. This phenomenon seems not to have been explored empirically. It is

372

Appendix B

likely that IRS assistance can help taxpayers understand how to take deductions properly and why it is in their interest to do so. Experiments with individuals or small regions might be undertaken to determine what kind of assistance works and whether such assistance changes either attitudes toward the IRS or behavior. SIMPLIFICATION

Tests of alternative tax forms lend themselves naturally to randomized experiments. That such forms may lead to better compliance is plausible and important if we judge by the work of the IRS and the states of Massachusetts and Pennsylvania. The individual taxpayer or couple are most pertinent for randomization. Formal randomized experiments on the topic seem not to have been conducted or published by state revenue offices. This is in contrast to the IRS's (1983) efforts to undertake and report on forms simplification. This early effort was not executed completely, but it did demonstrate the hazards of randomized field experiments with volunteer subjects—i.e., a 10 percent response rate of participation in tests of forms. ATTITUDES

The state of Massachusetts, among others, stresses the importance of changing attitudes in the interest of voluntary taxpayer compliance. This is an especially challenging matter since the relation between measured attitude and actual behavior is often not empirically explored. When the correlation is estimated in field surveys, it is often puny—e.g., correlations of less than .10 between work ethic attitudes and actual work behavior among youth involved in employment programs. Still, the matter is important to the extent that attitudes reflect actual behavior and attitudes are easier to measure. Experimentation seems feasible on a small scale, at least, if we use Schwartz-Orleans as a model. Indeed, the Schwartz-Orleans work suggests that remarkable attitude change and behavioral change resulted from their special treatments. (Attitudes were deduced from normative response to open-ended inquiry, not standardized-item questionnaires.) Experimental tests of attitude change devices seem most pertinent to those taxpayers whose behavior is least verifiable—e.g., the self-employed—but only if experiments on other groups support the notion that attitudes reflect actual behavior.

Experimental Designs

373

MANIPULABLE VARIABLES M O R E GENERALLY

Nagin's (1986) statement of the revenue office's missions and the Klepper-Nagin (1987) work make a nice little general guide to what variables might be manipulated in randomized tests. (a) Providing the means for compliance, as in simplifying or streamlining forms, and the information need to comply, as in revising instruction manuals, invites randomized tests of alternative forms, instruction manuals, and related client services. (b) Developing organization capacity to process returns invites experimental tests of alternative approaches to training revenue office staff members and of office procedures including communications, command and control, and so on. (c) Identifying noncompliers and bringing them into compliance invites field experiments that yield better identification schemes and various ways to encourage compliance. (d) Contributing to statutory changes that make the tax system better invites tests of experimental policies, procedures, or regulations that may eventually become law. All this is a bit too abstract to be very helpful, of course. But Nagin's model helps to build a speculative research agenda around the idea of systematic experiments in compliance.

References Aitken, S.S., and Bonneville, L. 1980 A General Taxpayer Opinion Survey. Prepared for Office of Planning and Research, Internal Revenue Service, March 1980, by C S R , Inc., Washington, D.C. (Contract No. TIR-79-2) Betsey, C., Hollister, R., and Papageorgiou, M., eds. 1985 Youth Employment and Training Programs: The TEDPA Tears. Washington, D.C.: National Academy Press. Bickman, L. 1985 Randomized field experiments in education. New Directions for Program Evaluation 28:39-54. Blumstein, A. 1983 Models for structuring taxpayer compliance. Pp. 159-172 in P. Sawicki, ed., Income Tax Compliance: A Report of the ABA Section on Taxation, Invitational Conference on Income Tax Compliance. Washington, D.C.: American Bar Association.

37+

Appendix B

Boruch, R.F. 1976 On common contentions about randomized experiments. Pp. 158-194 in G. V. Glass, ed., Evaluation Studies Review Annual, 3. Beverly Hills, Calif.: Sage Publications. Boruch, R.F., and Cecil, J.S. 1979 Assuring Confidentiality of Social Research Data. Philadelphia: University of Pennsylvania Press. 1983 (eds.) Solutions to Ethical and Legal Problems in Social Research. New York: Academic Press. Boruch, R.F., Dennis, M.L., and Greer, K.D. 1988 Lessons from the Rockefeller Foundation's experiments on the Minority Single Parent Program. Evaluation Review, in press. Boruch, R.F., McSweeny, A.J., and Soderstrom, J. 1978 Bibliography: randomized field experiments for program planning, development and evaluation. Evaluation Quarterly 4:655-696. Box, G.E.P., and Tiao, G.C. 1975 Intervention analysis with applications to economic and environmental problems. Journal of the American Statistical Association 70:70-92. Box, J.F. 1978 RA. Fisher: The Life of a Scientist. New York and Chichester: John Wiley & Sons. Brook, et al. 1984 The Effect of Coinsurance on the Health ofAdults: Results from the RAND Health Insurance Experiment. Santa Monica, CA: Rand Corp. Breger, M. 1983 Randomized social experiments and the law. Pp. 97—144 in R.F. Boruch and J.S. Cecil, eds., Solutions to Legal and Ethical Problems in Applied Social Research. New York: Academic Press. Campbell, D.T., and Boruch, R.F. 1975 Making the case for randomized assignment treatments by considering the alternatives: six ways in which quasi-experimental evaluations in compensatory education tend to underestimate effects. Pp. 195—296 in C.A. Bennett and A.A. Lumsdaine, eds., Central Issues in Social Program Evaluation. New York: Academic Press. Campbell, D.T., and Stanley, J.S. 1963 Experimental and Quasi-Experimental Designs for Research. Chicago: Rand McNally. Chelimsky, E., ed. 1985 Program Evaluation: Patterns and Directions. Washington, D.C.: American Society for Public Administration. Collins, J.F., and Elkin, I. 1985 Randomization in the NIMH treatment of depression collaborative research program. New Directions for Program Evaluation 28:27-38.

Experimental Designs

375

Conner, R.F. 1977 Selecting a control group: an analysis of the randomization process in twelve social reform programs. Evaluation Quarterly i(2):i95— 243-

Corsi, J.R. 1983 Randomization and consent in the New Mexico teleconferencing experiment: legal and ethical considerations. Pp. 159-171 in R. F. Boruch and J.S. Cecil, eds., Solutions to Ethical and Legal Problems in Social Research. New York: Academic Press. Corsi, J.R., and Hurley, T.L. 1979a Attitudes toward the use of the telephone in administrative hearings. Administrative Law Renew 31:247—283. 1979b Pilot study report on the use of the telephone in administrative fair hearings. Administrative Law Review 31:485~524. Dennis, M.L. 1988 Implementing Randomized Field Experiments: An Analysis of Criminal and Civil Justice Research. Ph.D. dissertation, Psychology Department, Northwestern University. Farrington, D.P. 1983 Randomized experiments on crime and justice. Crime and Justice: An Annual Review ofResearch 4:257-308. Farrington, D.P., Ohlin, L.E., and Wilson, J.Q. 1986 Understanding and Controlling Crime: Toward a New Research Strategy. New York: Springer-Verlag. Federal Judicial Center 1983 Social Experimentation and the Law. Washington, D.C.: Federal Judicial Center. Ferber, R., and Hirsch, W. Z. 1982 Social Experimentation and Economic Policy. Cambridge: Cambridge University Press. Fienberg, S.E., Martin, M., and Straf, M.L., eds. 1985 Sharing Research Data. Committee on National Statistics. Washington, D.C.: National Academy Press. Fienberg, S.E., Singer, B., and Tanur, J.M. 1985 Large-scale social experimentation in the United States. Chapter 12 in A.C. Atkinson and S. E. Fienberg, eds., A Celebration ofStatistics: ThelSI Centenary Volume. New York: Springer-Verlag. Fraker, T., and Maynard, R. 1984 The Use of Comparison Group Designs in Evaluations ofEmployment-Related Programs. Princeton, N.J.: Mathematica Policy Research. Freund, P.A. 1968 Legal frameworks for human experimentation. Daedalus 98:318. Friedman, L.M., Furberg, C.D. and DeMets 1985 Fundamentals of Clinical Trials. Littleton, Mass.: PSG Publishing, Inc. Gilbert, J.P., Light, R.M., and Mosteller, F. 1977 Assessing social innovations: an empirical base for policy. Pp. 185—242 in

376

Appendix B

W.B. Fairley and F. Mosteller, eds., Statistics and Public Policy. Reading, Mass.: Addison-Wesley. Glass, G.V. et al. 1975 Design and Analysis of Time Series Experiments. Boulder, Colorado: Colorado Associated University Press. Goldman, J. 1977 A randomization procedure for trickle-process evaluations. Evaluation Quarterly i(j):493-498. 1985 Negotiated solutions to overcoming impediments in a law related experiment. New Directions for Program Evaluation 28:63-72. Gordon, G., and Morse, E.V. 1975 Evaluation research. Annual Review of Sociology 1:339-361. Gray, B.H. 1975 Human Subjects in Medical Experimentation. New York: John Wiley & Sons. Gueron, J.M. 1985 The demonstration of work/welfare initiatives. New Directions in Program Evaluation 29:5-14. Gueron, J.M., and Nathan, R. 1985 The MDRC work/welfare project: Objectives, status, significance. Policy Studies Review 4(3). Hahn, G.J. 1984 Experimental design in the complex world. Technometrics 26(i):i9-3i. Hargrove, E.C. 1981 The bureaucratic politics of evaluation: A case study of the Department of Labor. In H.E. Freeman and M.A. Solomon, eds., Evaluation Studies Review Annual 6. Beverly Hills, Calif.: Sage Publications. Heckman, J.J., Hotz, V.J., and Dabos, M. 1987 Do we need experimental data to evaluate the impact of manpower training on earnings? Evaluation Review 11(4) 395-427. Hendricks, M. 1981 Service delivery assessment: Qualitative evaluations at the cabinet level. In N.L. Smith, ed., New Directions in Program Evaluation 12:5-24. Hillis, J.W., and Wortman, C.M. 1976 Some determinants of public acceptance of randomized control group experiments. Sociometry 39:91—96. Hillsman-Baker, S.B., and Rodriguez, O. 1979 Random time quota selection: An alternative to random selection in experimental evaluation. Pp. 185—186 in L. Sechrest et al., eds., Evaluation Studies Review Annual, 6. Beverly Hills, Calif.: Sage Publications. Hollister, R.G., Kemper, P., and Maynard, R.A. 1984 The National Supported Work Demonstration. Madison, Wis.: University of Wisconsin Press. Hornick, R.O. et al. 1973 Television and Educational Reform in El Salvador. Stanford, Calif.: Stanford University Institute for Communication Research.

Experimental Designs

377

Institute of Medicine 1985 Assessing Medical Technologies. Committee for Evaluating Medical Technologies in Clinical Use. Washington, D.C.: National Academy Press. Internal Revenue Service 1983 Long Range Tax Forms Simplifcation Study. Research Division, Tax Forms and Publications Division. Washington, D.C.: U.S. Department of the Treasury. Kinsey, K.A. 1984 Survey Data on Tax Compliance: A Compendium and Review. American Bar Foundation Tax Compliance Working Paper 84-1, December, 1984. American Bar Foundation, Chicago. 1987 Theories and models of tax evasion. Criminal Justice Abstracts 18.403: Revision of American Bar Foundation Tax Compliance Working Paper 84-2-, December 1984. American Bar Foundation, Chicago. Klepper, S., and Nagin, D. 1987 The Role of Tax Practitioners in Tax Compliance. Unpublished paper, Department of Statistics and School of Urban and Public Affairs, Carnegie-Mellon University. Lempert, R.O. 1984 From the editor. Law and Society 18(4) :503-504. Levine, R.J. 1981 Ethics and Regulation of Clinical Research. Baltimore and Munich: Urban and Schwarzenberg. Light, R.J., and Pillemer, D.B. 1984 Summing Up: The Science ofReviewing Research. Cambridge, Mass.: Harvard University Press. Lind, A. 1985 Randomized experiments in the federal courts. New Directionsfor Program Evaluation 28:73-80. Lipsey, M.W., Cordray, D.S., and Berger, D.E. 1981 Evaluation of a juvenile diversion program: Using multiple lines of evidence. Evaluation Review 5(3):283-3o6. Loftus, E. 1985 To file, perchance to cheat. Psychology Today, Aprils—39. Long, S.B. 1981 Social control in the civil law: the case of income tax enforcement. Pp. 185-214 in H.L. Ross, ed., Law and Deviance. Beverly Hills, Calif.: Sage Publications. Matt, G.E. 1988 Fraud and Error in Three Federal Programs: Aid to Families with Dependent Children, Food Stamps, and Individual Income Tax Returns. Ph.D. dissertation, Psychology Department, Northwestern University. McCleary, R., and Hay, R.A. 1980 Applied Time Series for the Social Sciences. Beverly Hills, Calif.: Sage Publications.

378

Appendix B

Meier, P. 1972 The biggest public health experiment ever: The 1954 field trial of the Salk poliomyelitis vaccine. In J.M. Tanur, F. Mosteller, W.H. Kruskal, R.F. Link, R.S. Pieters, and G. Rising, eds., Statistics: A Guide to the Unknown. San Francisco: Holden-Day. Mosteller, F., Gilbert, J.P., and McPeek, B. 1980 Reporting standards and research strategies for controlled trials: Agenda for the editor. Controlled Clinical Trials 1:37-58. Mosteller, F., and Weinstein, M.C. 1985 Towards evaluating the cost-effectiveness of medical and social experiments. In J.A. Hausman and D.A. Wise, eds., Social Experimentation. Chicago: University of Chicago Press. Nagin, D.S. 1986 The Tax Administrator's Model of Tax Compliance. Report prepared for the Conference on Tax Compliance Research, Padre Island, Texas, January i$—17,1986. Nelson, C.W. 1973 Applied Time-Series Analysis. San Francisco: Holden-Day. Nelson, R.L., and Hedrick, T.E. 1983 The statutory protection of confidentiality of research data: synthesis and evaluation. Pp. 213-236 in R.F. Boruch and J.S. Cecil, eds., Solutions to Ethical and Legal Problems in Social Research. New York: Academic Press. Perng, S.S. 1985 Accounts receivable treatments study. New Directions for Program Evaluation 28:55—62. Riecken, H.W. et al. 1974 Social Experimentation: A Method for Planning and Evaluating Social Programs. New York: Academic Press. Rivlin, R.M. 1971 Systematic Thinking for Social Action. Washington, D.C.: The Brookings Institution. Robins, P.K., Spiegelman, R.G., Weiner, S., and Bell, J.G. 1980 A Guaranteed Annual Income: Evidence from a Social Experiment. New York: Academic Press. Robins, P.K., and West, R.W. 1981 Labor Supply Response to the Seattle and Denver Income Maintenance Experiments, Menlo Park, Calif.: SRI International. Roos, L.L., Roos, N., and McKinley, B. 1977 Implementing randomization. Policy Analysis 3(4) 1547-559. Rossi, P.H., Berk, R.A., and Lenihan, K.J. 1980 Money, Work, and Crime: Experimental Evidence. New York: Academic Press. Rossi, P.H., and Freeman, H. 1981 Evaluation: A Primer. Beverly Hills, Calif.: Sage Publications. Rubin, D.B. 1977 Assignment of treatment group on the basis of a covariate. Journal of Educational Statistics 2:1—26.

Experimental Designs

379

Schueren, F. 1985 Methodologie Issues in Linkage of Multiple Data Bases. Prepared for the National Research Council Panel on Statistics for an Aging Population, September 13. Schultz, T.W. 1982 Distortions of economic research. Pp. 121-134 in W.H. Kruskal, ed., The Social Sciences: Their Nature and Uses. Chicago: University of Chicago Press. Schwartz, R.D., and Orleans, S. 1967 On legal sanctions. University of Chicago Law Review 34:274-300. Severy, L.J., and Whitaker, J.M. 1982 Juvenile diversion: An experimental analysis of effectiveness. Evaluation Review 6(6):753-774. Shadish, W.R., and Cook, T.D. in press Evaluation of social programs. In B.B. Wolman, ed., International Encyclopedia of Psychiatry, Psychology, Psychoanalysis, and Neurology (First Progress Volume). New York: Aesculapius Press. Sherman, L.W., and Berk, R.A. 1984 The specific deterrent effects of arrest for domestic assault. American Sociological Review 49:261—272. 1985 Randomized experiments in police research. New Directions for Program Evaluation 28:15-25. St. Pierre, R.G., and Rezmovic, V. 1982 An overview of the Nebraska Nutrition and Education Training Program evaluation. Journal ofNutritional Education 14(2) :6i-66. St. Pierre, R.G., Cook, T.D., and Straw, R. 1982 Evaluation of the Nebraska Nutrition and Education Training Program Evaluation News 3(1) :67-69. Trochim, W.M.K. 1984 Research Design for Program Evaluation: The Regression Discontinuity Approach. Beverly Hills, Calif.: Sage Publications. Witte, A.D. 1987 Tax compliance research. Pp. 101-109 in Proceedings of the American Statistical Association: Survey Research Methods Section. Washington, D.C.: American Statistical Association.

Appendix C: Symposium on Taxpayer Compliance Research

National Academy of Sciences/National Research Council Commission on Behavioral and Social Sciences and Education Committee on Research on Law Enforcement and the Administration of Justice Panel on Taxpayer Compliance Research AGENDA SYMPOSIUM ON TAXPAYER COMPLIANCE RESEARCH January 15-17,1986 Conference Room A Padre Island Hilton Resort South Padre Island, Texas

Wednesday, January is 8:30 a.m. CONTINENTAL BREAKFAST 9:00 a.m.

Welcome and Introductions Ann Witte, Symposium Chair

9:15

Roundtable I. What we need to know: Political and administrative perspectives. (Panel Report Outline II.A, V.) Jerome Kurtz, Chair Discussants: Ira Jackson (Steven Klepper/Daniel Nagin paper) Paul McDaniel (John Scholz paper) John Wedick (Synthesis) Eugene Bardach (Synthesis)

Symposium on Taxpayer Compliance Research

381

Authors: Steven Klepper and Daniel Nagin 'The Tax Administrator's Model of Compliance" John Scholz "Political Context of Tax Administration" Panel Commentators: Walter Blum Alfred Blumstein 12:15 p.m.

LUNCH

2:15 p.m.

Roundtable II. The individual incentive model: Costs and benefits of compliance. (Outline section III.A.i—3) Harvey Galper, Chair Discussants: Daniel Nagin (Robert Kagan paper) Steven Klepper (Joel Slemrod paper) Thomas Coleman (Synthesis) Paul McDaniel (Synthesis) Authors: Robert Kagan "Visibility of Violations and Tax Compliance" Joel Slemrod "Complexity, Compliance Costs, and Tax Evasion" Panel Commentators: Walter Blum Stewart Macaulay

5:15 p.m.

A D J O U R N FOR T H E DAY

Thursday, January 16 8:30 a.m.

CONTINENTAL BREAKFAST

9:00 a.m.

Roundtable III. Extending the individual incentive model. (Outline section III.A.4.) Richard Schwartz, Chair Discussants: Daniel Rubinfeld (Howard Margolis and Suzanne Scotchmer papers)

382

Appendix C Richard Nisbett (John Carroll paper) Ira Jackson (Synthesis) Authors: John Carroll "Role of Uncertainty in Compliance Decision-Making" Howard Margolis "Taxpayer Noncompliance and the Problem of Collective Action" Suzanne Scotchmer "Extensions of Theoretical Modeling of Tax Compliance" Panel Commentators: Jerry Green Jerome Kurtz 12:00 p.m.

LUNCH

2:00 p.m.

Roundtable IV. Social influences. (Outline section III.B.) Richard Lempert, Chair Discussants: Felice Levine (Robert Kidder/Craig McEwen paper) John Carroll (Robert Cialdini paper) Howard Erlanger (Third Parties) Roger Plate (Synthesis) Authors: Robert Kidder and Craig McEwen "Normative Pluralism and Compliance Research" Robert Cialdini "Social Motivation to Comply: Intrinsic Rewards and Social Influences" Panel Commentators: Sidney Davidson Richard Nisbett Barbara Yngvesson

5:00 p.m.

A D J O U R N FOR T H E D A Y

6:00 p.m.

Reception and Dinner Conference Room B

Symposium on Taxpayer Compliance Research

383

Friday, January 17 8:30 a.m.

CONTINENTAL BREAKFAST

9:00 a.m.

Roundtable V. Future Research. (Outline section IV, V.) Alfred Blumstein, Chair Discussants: Daniel Rubinfeld (Robert Boruch/Peter Schmidt papers) Harold Grasmick (Seymour Sudman paper) Felice Levine (Synthesis) Fred Goldberg (Synthesis) Authors: Robert Boruch "Experiments and Quasi-Experiments in Context of Tax Compliance Research" Peter Schmidt "Statistical Considerations in Modeling Taxpayer Compliance" Seymour Sudman "Observational Techniques for Tax Compliance Measurement" Panel Commentators: Jan Kmenta David Linowes

12:00 p.m.

LUNCH

1:00 p.m.

A D J O U R N SYMPOSIUM—Authors and invited participants depart for airport

Appendix D: Panel on Taxpayer Compliance Research

ANN DRYDEN WITTE (Chair), Department of Economics, Wellesley College, and National Bureau of Economic Research, Cambridge, Massachusetts EUGENE S. BARDACH, Graduate School of Public Policy, University of California, Berkeley WALTER J. BLUM, School of Law, University of Chicago ALFRED BLUMSTEIN, School of Urban and Public Affairs, Carnegie-Mellon University SIDNEY DAVIDSON, Graduate School of Business, University of Chicago HARVEY GALPER, Peat Marwick Main & Co., Washington, D.C. JERRY R. GREEN, Department of Economics, Harvard University JAN KMENTA, Department of Economics, University of Michigan JEROME KURTZ, Paul, Weiss, Rifkind, Wharton and Garrison, Washington, D.C. RICHARD O. LEMPERT, School of Law, University of Michigan DAVID F. LINOWES, Institute of Government and Public Affairs, School of Public Policy, University of Illinois STEWART MACAULAY, School of Law, University of Wisconsin RICHARD E. NISBETT, Institute for Social Research, University of Michigan JOHN W. PAYNE, Fuqua School of Business, Duke University RICHARD D. SCHWARTZ, School of Law, Syracuse University BARBARA YNGVESSON, School of Social Science, Hampshire College JEFFREY A. ROTH, Study Director J O H N T . SCHOLZ, Senior Research Associate GAYLENE J. DUMOUCHEL, Administrative Secretary TERESA E. WILLIAMS, Administrative Secretary

Panel on Taxpayer Compliance Research

385

Biographical Sketches A N N D R Y D E N Wl l ' l E is professor of economics at Wellesley College and research associate at the National Bureau of Economic Research. In her research, she applies microeconomic theory and econometric techniques to the study of various social issues, including tax compliance, law enforcement, crime, housing markets, and day care. She has authored or edited seven books including Predicting Recidivism Using Survival Models; Advances in Applied Micro-Economics; An Economic Analysis ofCrime andJustice; and Beating the System: The Underground Economy. Witte has served as chair of the American Economics Association's Census Advisory Commission and has been a member of numerous professional and advisory committees, including the panel's parent Committee on Research on Law Enforcement and the Administration of Justice and the American Statistical Association's Committee on Law and Justice Statistics. She is a fellow of the American Society of Criminology, a trustee of the Police Foundation and the Law and Society Association, and an associate editor of a number of professional journals. She received a B.A. from the University of Florida, an M.A. from Columbia University, and a Ph.D. in economics from North Carolina State University. E U G E N E S. B A R D A C H is professor of public policy at the Graduate School of Public Policy, University of California, Berkeley. He has written on policy and program implementation and especially, in recent years, on the implementation of health and safety regulation by all levels of government. His books include The Implementation Game: What Happens After a Bill Becomes a Law; Going by the Book: The Problem ofRegulatory Unreasonableness (with Robert A. Kagan); and The Skill Factor in Politics: Repealing the Mental Commitment Laws in California. He is currently writing a book on ethics and public policy. He received a B.A. from Columbia University and a Ph.D. in political science from the University of California, Berkeley. W A L T E R J. B L U M is the Edward H. Levi distinguished service professor at the University of Chicago Law School. He has served as trustee of the College Retirement Equities Fund and as consultant to the American Law Institute Federal Income Tax Project. From i960 to 1966 he served on the Advisory Committee for Studies of Government Finance at the Brookings Institution and was a member of the Steering Committee of the Administrative Conference Project on the U.S. Internal Revenue Service (19741975). Since 1947 he has been a member of the Planning Committee of the

386 Appendix D University of Chicago Law School Federal Tax Conference and since 1948 has been legal counsel to the Bulletin ofthe Atomic Scientists. He is a member of the American Bar Association, the Chicago Bar Association, the American Law Institute, the Chicago Federal Tax Forum, and the American Academy of Arts and Sciences. He is the author of The Uneasy Case for Progressive Taxation (with Harry Kalven, Jr.); Public Law Perspectives on a Private Law Problem (with Harry Kalven, JrMaterials on Reorganization, Recapitalization and Insolvency (with Stanley A. Kaplan); and Readjustments and Reorganizations (with Stanley A. Kaplan); as well as numerous articles in the fields of federal taxation, insurance, corporate finance, and bankruptcy. ALFRED BLUMSTEIN is dean and J. Erik Jonsson Professor of urban systems and operations research at the School of Urban and Public Affairs of Carnegie-Mellon University. He has had extensive experience in both research and policy with the criminal justice system since serving the President's Commission on Law Enforcement and Administration of Justice in 1966—1967 as director of its Task Force on Science and Technology. His research has covered many aspects of the operation of the criminal justice system, with special attention to prison populations and sentencing, and has involved extensive research on criminal careers. Blumstein was a member of the panel's parent Committee on Research on Law Enforcement and Administration of Justice from its founding in 1975 until 1986. He served as chair of the committee between 1979 and 1984 and has chaired the committee's panels on research on deterrent and incapacitative effects, on sentencing research, and on research on criminal careers. Blumstein has served since 1979 as chairman of the Pennsylvania Commission on Crime and Delinquency, the state criminal justice planning agency for Pennsylvania. He has also been a member of the Pennsylvania Commission on Sentencing since 1986. His degrees from Cornell University include a baccalaureate in engineering physics and a Ph.D. in operations research. He was president of the Operations Research Society of America in 19771978 and was awarded its Kimball Medal for service to the profession and society in 1985. He is currently the president of the Institute of Management Sciences and a fellow of the American Society of Criminology and was the 1987 recipient of the society's Sutherland Award for research contributions. SIDNEY DAVIDSON is the Arthur Young Distinguished Service Professor of accounting at the University of Chicago. He is a certified public

Panel on Taxpayer Compliance Research

387

accountant and holds B.A., M.B.A., and Ph.D. (in business administration) degrees from the University of Michigan. He is the author or editor of several books and many articles on accounting, economics and finance. He was president of the American Accounting Association in 1968—1969 and vice president of the American Institute of CPAs in 1986—1987. He was elected to the Accounting Hall of Fame in 1983. He has served as a consultant to the U.S. Department of the Treasury and the Securities and Exchange Commission. HARVEY GALPER is a principal in the accounting firm of Peat Marwick Main & Co., resident in its Washington, D.C., office. His prior positions include senior fellow at the Brookings Institution (1982—1987), senior public finance resident at the Advisory Commission on Intergovernmental Relations (1981-1982), and director of the Office of Tax Analysis of the U.S. Treasury Department (1976-1981), from which he received a meritorious service award. He has also served on the staff of the Urban Institute and has taught at Dartmouth College, the University of California, Berkeley, Yale University, and the Georgetown University Law Center. He has authored or coauthored more than forty articles in public finance, wrote Assessing Tax Reform with Henry Aaron in 1985, and edited Uneasy Compromise: Problems of a Hybrid Income-Consumption Tax with Henry Aaron and Joseph Pechman in 1988. He is a past member of the board of directors of the National Tax Association-Tax Institute of America and currently serves on the editorial board of the National Tax Journal and on the advisory group of the commissioner of the Internal Revenue Service. He received a B.A. from Dartmouth College and M.A. and Ph.D. degrees from Yale University. JERRY GREEN is the David A. Wells Professor of political economy at Harvard University. He has served as chairman of the Department of Economics at Harvard and as chairman of the National Science Foundation's Information Science Advisory Panel. He is a research associate of the National Bureau of Economic Research, a fellow of the Econometric Society, and an overseas fellow of Churchill College, Cambridge University. He has written over seventy articles published in economics journals and has edited several collections of papers. His book, coauthored with Jean-Jacques Laffont, Incentives in Public Decision Making, was published in 1979. He has held a Woodrow Wilson dissertation fellowship (1969— 1970), and a National Science Foundation postdoctoral fellowship (1971—1972), has been a fellow of the Center of Advanced Study in

388

Appendix D

Behavioral Science (1980-1981), and is currently a John Simon Guggenheim memorial fellow. He has been the recipient of the J. K. Galbraith Prize for teaching in economics (1980) and the distinguished alumni award of the University of Rochester (1984). He is founder and editor of Economic Letters. He received a B.A. and a Ph.D. in economics from the University of Rochester. J A N K M E N T A is professor of economics and statistics at the University of Michigan, specializing in econometric models and methods. His research has involved a variety of theoretical and applied econometric problems. He is a fellow of the American Statistical Association and of the Econometric Society and associate editor of the Journal of the American Statistical Association and of the Review of Economics and Statistics. His publications include Elements of Econometrics; Evaluation of Econometric Models (coedited with James B. Ramsey); and Large Scale MacroEconometric Models: Theory and Practice (coedited with James B. Ramsey). He received a B.Ec. degree from the University of Sydney, Australia, and M.A. and Ph.D. degrees in economics and statistics from Stanford University. J E R O M E K U R T Z is a partner in the law firm of Paul, Weiss, Rifkind, Wharton & Garrison, resident in its Washington, D.C., office. He has served in the U.S. Department of the Treasury as commissioner of internal revenue (1977—1980) and as tax legislative counsel (1966—1968); he received that department's Alexander Hamilton Award (1980) and its Exceptional Service Award (1968). He has taught tax law and policy as a lecturer at Villanova University (1964—1965) and the University of Pennsylvania Law School (1969-1974) and as visiting professor of law at Harvard University (1975-1976). He served as chairman of the tax section of the Philadelphia Bar Association (1975—1976), a member of the executive committee of the tax section of the New York State Bar Association (1981-1982), chairman of the special committee on tax shelters of the tax section of the American Bar Association (1982-1984), and president of the Center for Inter-American Tax Administration (1980). He is a member of the Federal Income Tax Project Advisory Group of the American Law Institute and a fellow of the American College of Tax Counsel. He received a B.S. from Temple University and an LL.B. from Harvard University. R I C H A R D O. L E M P E R T is professor of law and sociology at the University of Michigan. He received an A.B. degree from Oberlin College,

Panel on Taxpayer Compliance Research

389

a J.D. from the University of Michigan Law School, and a Ph.D. in sociology from the University of Michigan. He is chair of the Committee on Research on Law Enforcement and the Administration of Justice. He has served as editor of the Law and Society Review and as a member of the executive committee of the Law and Society Association. He is coauthor of A Modern Approach to Evidence and An Invitation to Law and Social Sciences. D A V I D F. LINOWES is professor of political economy and public policy and Boeschenstein Professor Emeritus at the University of Illinois. He is also senior advisor to the Institute of Government and Public Affairs. Linowes is chairman of the President's Commission on Privatization. He served as chairman of the President's Commission on the Nation's Energy Resources and from 1975 to 1977 as chairman of the U.S. Privacy Protection Commission. He was a founding partner of Laventhol and Horwath, a worldwide management consulting firm. He headed economic development missions for the U.S. Department of State and the United Nations to Turkey, India, Greece, Pakistan, and Iran in the late 1960s and early 1970s. He is the author of Managing Growth Through Acquisition, Strategies for Survival, The Corporate Conscience, and The Privacy Crisis in Our Time. Linowes is a member of the Council on Foreign Relations. STEWART M A C A U L A Y is Malcolm Pitman Sharp Professor of the University of Wisconsin. He is a member of the Commission on Behavioral and Social Sciences and Education (CBASSE). He was president of the Law and Society Association and associate editor of The Law & Society Review. He is coauthor (with Lawrence Friedman) of Law and the Behavioral Sciences, one of the first collections of teaching materials in the field. His "Non-Contractual Relations and Business: A Preliminary Study," published in 1963, is still widely cited and reprinted. More recent publications include "Private Government," "Law and the Behavioral Sciences: Is There Any There There?" and "Images of Law in Everyday Life: The Lessons of School, Entertainment, and Spectator Sports." He received A.B. and LL.B. degrees from Stanford University. R I C H A R D E. NISBETT is professor of psychology at the University of Michigan and program director at the Institute for Social Research. His research interests include judgment and decision making, inductive reasoning, and attitude change. He is coauthor or editor of seven books, including Human Inference (with L. Ross) and Induction (with J. Holland, K. Holyoak, and P. Thagard). He is the former director of the cognitive science

390

Appendix D

program of the University of Michigan. In 1982 he was awarded the Donald Campbell Prize for distinguished research in social psychology by the American Psychological Association. He received a Ph.D. in social psychology from Columbia University. JOHN W. PAYNE is professor of business administration at Duke University. He is director of the Center for Decision Studies at the Fuqua School of Business and area coordinator for the management and organizational behavior faculty. He has published forty articles and book chapters dealing with the psychology of decision making. He is an associate editor of several journals in management science and decision research. He received B.A., M.A., and Ph.D. degrees in psychology from the University of California, Irvine, and was a postdoctoral fellow in cognitive psychology at Carnegie-Mellon University. JEFFREY A. ROTH, who served as the panel's study director, is study director for the Committee on Research on Law Enforcement and the Administration of Justice. His interest is in the policy uses of social research, especially in the areas of taxpayer compliance, criminal careers, and pretrial release. He is a member of the American Society of Criminology, the Law & Society Association, the American Economic Association, and the American Statistical Association. He received B.A., M.A., and Ph.D. degrees in economics from Michigan State University. RICHARD D. SCHWARTZ is Ernest I. White Professor of law at the College of Law and professor of sociology and social science at the Maxwell School of Syracuse University. He is the former dean and provost of the Faculty of Law and Jurisprudence at the State University of New York, Buffalo, and was founding editor of the Law and Society Review. He was a member of the Committee on Research on Law Enforcement and the Administration of Justice. He is interested particularly in the sociology of law, criminal law and society, and administrative regulation; a current focus is public participation in environmental protection proceedings. He received B.A. and Ph.D. degrees in sociology from Yale University. JOHN T. SCHOLZ, research associate for the panel, is associate professor and director of the graduate program in political science at the State University of New York, Stony Brook. His primary research and publications have focused on enforcement and compliance aspects of regula-

Panel on Taxpayer Compliance Research

391

tory and tax policies, particularly the Occupational Safety and Health Administration and the Internal Revenue Service. He has also written on other aspects of regulatory policy, and co-authored Nepal: Profile of a Himalayan Kingdom. He received a B.A. in government from Harvard University and M.A. and Ph.D. degrees in political science and an M.S. in environmental economics from the University of California, Berkeley. BARBARA YNGVESSON is professor of anthropology in the School of Social Science at Hampshire College. She has written widely on informal social control, dispute processing, and the interpretive theory of law. Her research has been carried out in Sweden and in the United States and has focused on the interplay of local and official understandings and practices. She is currently completing a book on the negotiation and transformation of local conflicts by officials and citizens in an American criminal court and the implications of this negotiation for understanding of the social construction of law.

Appendix E: Committee on Law Enforcement and the Administration of Justice, 1987-1988

RICHARD LEMPERT (Chair), School of Law, University of Michigan ALBERT J. REISS, JR. (Vice Chair), Department of Sociology, Yale University ANTHONY V. BOUZA, Chief of Police, Minneapolis Police Department JONATHAN D. CASPER, Department of Political Science, Northwestern University, and American Bar Foundation, Chicago, Illinois JACQUELINE COHEN, School of Urban and Public Affairs, Carnegie-Mellon University PHILIP COOK, Institute of Public Policy, Duke University SHARI S. DIAMOND, Department of Psychology, University of Illinois at Chicago, and American Bar Foundation, Chicago, Illinois DAVID P. FARRINGTON, Institute of Criminology, Cambridge University, England ROBERT KAGAN, Department of Political Science, University of California, Berkeley MARK H. MOORE, Kennedy School of Government, Harvard University JOHN ROLPH, The Rand Corporation, Santa Monica, California KURT L. SCHMOKE, Mayor, Baltimore, Maryland JAMES F. SHORT, JR., Social Research Center, Washington State University PATRICIA MCGOWAN WALD, U.S. Court of Appeals for the District of Columbia Circuit

Committee on Law Enforcement

393

STANTON WHEELER, Yale Law School, Yale University BARBARA YNGVESSON, School of Social Science, Hampshire College ANN DRYDEN WITTE (ex officio), Chair, Panel on Taxpayer Compliance Research; Department of Economics, Wellesley College, and National Bureau of Economic Research, Cambridge, Mass. JEFFREY A. ROTH, Study Director TERESA E. WILLIAMS, Administrative Secretary

Index

A b t Associates, 358-359, 360 Accountants. See Tax practitioners Accounts Receivable Treatment Study (IRS), 241-242, 342, 356, 363 Adjusted gross income ( A G I ) , 44, 105, 319 A g e , 81; compliance levels and, 8, 133-135, 329; and social sanctions, 157; and tax practitioner use, 174 Agency Practice Act o f 1965, 35 Aggregate cross-sectional data base. See

81, 97-99, 102-106; in research, viii, 23-24, 93-96, 102-106; revenue yield from, 30-31; selection strategies, 83-86; taxpayers and, 71-72, 90, 327-328; T C M P , 2, 49, 58, 66-68, 308, 309, 316, 317. See also Enforcement, tax law Banks, 26 Barter, 92

Project 778 Aggregate unreported taxable personal income,/^. 42 Aguayo v. Richardson, 243, 348 Albany, N . Y . , 167 Alimony income, 29, 60 Allingham, Michael G., 80-82, 83 Altruistic acts, 124

Behavioral decision theory, 152

Ambiguity, 22, 26, 27, 81, 88; perceptions of, 128-129; tax practitioners and, 175. See also Complexity American Bar Association ( A B A ) , 169, 252; Commission on Taxpayer Compliance, 20; ethical standards, 193-194; and research, 17; Tax Section, 16 American Bar Foundation, 68, 190, 195 American Civil Liberties Union, 357 American Institute o f Certified Public Accountants ( A I C P A ) , 16, 176, 193-194

Bureau o f Economic Analysis ( B E A ) ,

Boruch, Robert F., ix, 339 Bounded rationality theory, 155-156, 160 Bouza, Anthony V . , 392 Brackets, tax, 182 Brandeis, Louis, 348 Brookings Institution, 341 Budget deficits, 16, 47 44-45 Burke, Edmund, 339 Business expenses: extent o f noncompliance, 43,52,54,59, 61, 64, 65; perceptions of, 132 Calibration experiments, 356 California, 159 California Welfare Rights Organization

v.

Richardson, 348 Campbell, Donald T., 341, 368

American Medical Association, 169 Anthropologists, 77 Arthur Y o u n g Foundation, 17, 252 Attorneys, 35. See also Tax practitioners Audit-based compliance measure, 222-223,

Capital gains income: extent o f

224, 231-232 Audits: cost of, 30, 31; D I F and, 186; interactive strategies, 83, 84-86; multiperiod dynamic model, 83, 86-87; perceptions of, 54, 74-75; random, 83-84; rate of, viii, 6, 30, 72,

Cash income, 4 4 - 4 5 , 107

noncompliance, 26, 52, 60, 63, 107-108, 115; overreporting, 26,52, 60, 63, 70 n. 9 Carroll, John, 152 Casualty losses, 61 Catfish farmers, 167 Censored normal regression model, 318-319 Census Bureau, 45, 46, 214-215, 218, 228, 235, 3+3, 350

Index Center for Tax Compliance Research, 263 Certified public accountants, 35, 176. See also Tax practitioners Charitable contributions, 99 Checking accounts, 44 Child care credit, 52, 61, 63 Cialdini, Robert B., 153-154 Circular 230 (Treasury Department), 35, 193-194

Clotfelter, Charles T., 307, 319 Clustering at zero, 317-327, 335 Cognitive decision making, 182-183 Colombia, 360 Commitment/consistency principle, 153-155, 204-205 Complexity, 26, 27, 81, 88, 148; occupation and, 219; and overreporting, 22-23, 59-63; of tax forms, 32-33, 118, 161, 197; taxpayers and, 32, 49, 53-54, 128-129, 142-144, 156, 174 Compliance: audit-based measure of, 222-223, 224; defined, 1-2, 21-23, 69 n. 3; return-based measure of, 99, 232; survey-based measure of, 222-227 Compliance costs, 3-4, 27, 72-73, 81, 87, 88, 148; reducing, 32-34; research on, 118; sources of, 89-90, 128-129 Compliance Measurement Handbook (1RS), 342

Compliance Research Advisory Group, 14, 209-210, 216, 217, 219, 229, 248, 256, 260-262 Computers, 37; arithmetic error checking, 48, 58; audit targeting with, 186; in enforcement, 17, 28, 36, 72, 98, 199; mailings, 69 n. 1; in research, 106, 199—202; and visibility, 29, 107, 228 Construction contractors, 31 Contingent process decision rule, 153 Continuous Longitudinal Manpower Survey, 364, 365 Contraceptive use, 166 Controlled Clinical Trials, 341 Corporate taxes, 19, 41, 42, 43, 184 Corrigan, Truax v., 348 Cragg, John G., 319-320, 325 Crane v. Mathews, 353 Credit cards, 44 Credits, tax, 20, 52, 54, 61, 63, 109 Crime, white-collar, 166

395

Crime Control Act of 1973, 352 Crime Control Institute, 358 Criminality, predictors of, 122 Cross-sectional studies, 350, 364 Currency, value of, 44 Current Population Survey, 365

Decision making, no, 152, 182-183, 234, 253; rules and heuristics, 99-100, 152-153, 182-183 Deductions, 20, 134; medical expense, 29-30, 52, 61, 63, 120, 125—126; misreported, 25-26, 41, 43, 45, 50-51, 56-58, 64; noncomplier prevalence in, 52; perceptions of, 75, 132; in statistical analysis, 312-313; TRA86 and, 16-17, 182, 184; visibility of, 29, 109; V R P of, 61, 64, 65 Deficit Reduction Act of 1984, viii Deliberate noncompliance, 308-310, 312, 315-316, 323-324 Demographics, 8, 81, 227; relation to compliance, 133-137; in statistical analysis, 330-331; in surveys, 334 Denmark, 370 Detection, probability of, 5-6, 27, 81, 83, 87, 88, 90, 139 n. 6; cost of increasing, hi—112; and deterrence, 96-97, 100—101; research on, 83, 96-110, 152; visibility and, 71—72, 106-110. See also Enforcement, tax law Deterministic frontier models, 308, 312 Deterrence, 3, 6, 148, 253; audits and, 72, 93-96, 102-106; and detection probability, 96-97, 100-101; general, 28, 102-106; and IRS activity, 212; research on, 78-79; and social sanctions, 113; specific, 93-96 Diamond, Shari S., 243 Discriminant Index Function (DIF), 30, 36, 66-69, 98-99; and audit patterns, 186; protecting secrecy of, 239-240; selection system, 84, 85; TCMP cycles and, 214, 215, 216; timeliness of, 208 Dividend income, 26, 28, 52, 53,59, 60, 64, 65, 107-108 Domestic servants, 168 Drug Abuse Office and Treatment Act, 352 Drugs, illegal, 167

396

Index

Economists, 77, 89 Education: correlation with compliance, 133, 134, 137-138; and nonfiling, 118; and tax practitioner use, 174 Elderly, 33, 169 El Salvador, 359 Employer withholding, 43 Enforcement, tax law: attitudes toward, 81, 130-133; audit component of, 30; computers and, 17, 28, 36, 48,58, 69 n. 1, 72, 98, 186, 199; costs of, in—112, 177; effects on taxpayers, 102-103, 168-169, 171, 186-188, 208, 227; ineffectiveness of, 1, 48, 139 n. 6; security of, 239-240, 257-258; T C M P as, 67. See also Audits; Penalties Equity, 81; perceptions of, 120-121, 127-128, 154, 184-186, 253; research and, 171 Estate income, 108 Estimated tax, 142 Ethics, practitioner, 193-194, 195-197 Ethnographic studies, 9, 221 Evaluation Review, 341 Evasion, tax, 20; defined, 22; perceptions of, 119—120 Exchange equity, 127-128 Excise taxes, 18 "Exemplary practitioner" program, 195 Exemptions, 20, 183 Expected utility, 80-82, 83, 89, 139 n. 1, 175 Experiential hypotheses, 1 0 1 - 1 0 2 Experimentation and the Law (Federal Judicial Center), 341 Fairness. See Equity Farm income, 52, 53, 60—64, 107-108, 165 Federal Judicial Center, 242, 341 Field experiments, 230—232 Fields, social, 165-171 Fishing industry, 167 Fixed-effects model, 330, 331 Food and Drug Administration, 352—353 Formal sanctions. See Penalties Form 1040, 1040A, 1 0 4 0 E Z . See Tax forms Form 1099. See Information returns; Withholding Fourteenth Amendment, 348 France, 130 Frontier production functions, 308-309, 312, 315

Game theory, 254 General Accounting Office (GAO), 16, 139 n. 6; on I R S , 33, 34, 176, 186; and research, 254-255 General Enforcement Program, 31 General Taxpayer Opinion Survey, 343 Georgia, 345 Germany, 130 Government: attitudes toward, 73, 81, 126; legitimacy of, 148 Great Depression, 163-164 Great Society, 357 Habitual compliance, 151 Halo effects, 196 Hastings Center, 353-354 Hawthorne effect, 244 Heckman, James J., 320, 325 Heterogeneity, 333 Holmes, Oliver Wendell, 348 Horizontal equity, 127-128 H & R Block, 118 Income: adjustments to, 20, 50-51, 65; audits and, 30; correlation with compliance, 7, 72, 116-117, 133, 134, 137-138; forms of, 87-88, 107, 108, 218; illegal, 19, 43, 44, 47; interest and dividend, 26, 28, 52, 53, 59, 60, 63, 64, 65, 99, 107-108; in microeconomic research, 83-84; and moral commitment, 367; partnership, 52, 108; pension, 60; personal, unreported taxable, 41 42; prevalence of noncompliance, 50—51; reporting, perceptions of, 125-126, 132; secondary, 218; sources of, 217, 221, 226; in statistical analysis, 312-313; and tax practitioner use, 174; unreported, 2, 24, 25-26, 41, 42, 45, 50-51, 3I2-3I3See also Visibility of income Income tax noncompliance gap, 41, fig. 43,

46-47

Independent contractors, 28 Individual income tax reporting gap, 24-25, 41-42, 47

Individual retirement accounts ( I R A ) , 28, 52, 54, 59, 61 Informal social sanctions, 7 - 8 , 73, 81, 91-92, 157; research on, 77—80, 112-113

Index Informal suppliers, 46, 167 Information returns, 27, 28-29, 45~+6, 48, 87, 228; legislation, 16-17, 106; in TCMP audits, 67; and visibility, 26, 71, 99, 107; and VRP, 59. See also Withholding Information Returns Program (IRP), 28-29, 30-31, 197, 200 Institute of Medicine, 355, 365 Institutional review boards, 235-236, 352-354

Interactive audit strategies, 83, 84-86 Interagency Human Subjects Coordinating Committee, 241, 354 Interest income: noncompliance amounts, 107-108; noncomplier prevalence in, 52; visibility of, 99; VRP, 28, 60, 63, 64, 65 Interest penalty, 61 Intergovernmental Personnel Agreements, 259

Intermediate transactions, 41 Internal Revenue Code, 2, 20, 21; perceptions of, 127-129, 154, 184-186; politics and, 187-188; in research, viii; TRA86 and, 186 Internal Revenue Service (1RS): budget, 30, 37, 104, 105-106; and compliance costs, 32-34; compliance strategies of, 27; and confidentiality of tax information, 148, 327, 350; Criminal Investigation Division, 32; data, 5, 18, 48, 236-241, 257-262; Examination Division, 216; master file, 68; Office of Planning and Research, 343; and penalties, 111-112; perceptions of, 37-38, 74-76, 81, 92, 104, 126, 129, 142-144, 167; processing problems of 1985, 189; Public Affairs Division, 34; and research, external, 14, 37-39, 230-232, 241-245, 248, 251-265, 342, 356; Research Division, ix, 14, 27, 35-36, 59, 213, 236, 241, 253, 256-257, 2 59, 363-364; sensitive data, 209, 213, 234-236; Statistics of Income Division, 236, 238; Strategic Plan, 197; Tax Forms and Publications Division, 32-33; tax gap estimates, vii, 1, 41-42, 45-46, 47; taxpayer contacts, 10, n-12, 33 - 34, '60, 161, 173, 183-184, 197-206,

397

371-372; and tax practitioners, 27, 33, 34-35, 175-176, 203; Treasury Circular 230, 35, 193-194; uncertainty of, 90. See also Audits; Enforcement, tax law; Project 778; Survey of Individual Returns Filed; Taxpayer Compliance Measurement Program Internal Revenue Service, Long v., 68, 239 Investment credit, 52, 54, 61, 63 Israel, 84 Italy, 130 Iterative experiments, 366 Jackson, Ira, 371 Kagan, Robert A., 107 Kidder, Robert, 219 Klepper, Steven, 90, 91, 121 Kohlberg, Lawrence, 156 Kruskal, William, 341 Laboratory experiments, 9, 232-234, 246 n. 14, 3 4 4

Law: attitudes toward, 125-126; regulatory, 147-148. See also Internal Revenue Code Lawyers, tax. See Tax practitioners Lewis, Alan, 119, 123 Liebmann, New York State he v., 348 Locked-box technique, 245 n. 7, 343 Long v. Internal Revenue Service (1979), 68, 239 Longitudinal studies, 364-365; data files, 240; protecting privacy in, 350; of taxpayers and practitioners, 192-193 Long Range Tax Forms Simplification Study (IRS), 344-345, 348, 363-364, .367

McCarthy, Charles, 354 McEwen, Craig, 219 Machiavelli, 118 Maiman, Richard J., 129, 169 Manpower Development Research Corporation (MDRC), 357, 358, 359 Marijuana use, 141 n. 19 Massachusetts, 370, 371, 372 Massachusetts Board of Retirement v. Murgla, 243

398

Index

Mass media contacts, IRS, 197, 198, 204-206 Mathematica Policy Research, 358 Mathews, Crane v., 353 Maximum likelihood estimate (MLE), 310-312, 314, 319, 326 Measurement error, 323-326 Medical expense deductions, 29—30, 120, 125-126; noncomplier prevalence in, 52; VRP, 61, 63 Microeconomic theories, 7, 72, 80-90, 152-153 Minneapolis Domestic Violence experiment, 345, 357, 361 Misclassifications, 21, 28 Monte Carlo experiment, 311 Moral commitment, 75-76, 81, 91-92, 124; income and, 367; and perceptions of equity, 128; research and, 8, 73, 77-80, 113, 118-133, 342-343 Moral development theory, 156 Mortgage interest, 26, 52, 54, 61 Mosteller, Frederick, 341 Multiperiod dynamic models, 81, 83, 86—87 Murgla, Massachusetts Board of Retirement v., 243 Nagin, Daniel, 90, 91, 121 Nash equilibrium model, 85 National Academy of Sciences, ix, 1, 263 National Center for Health Statistics, 343 National income and product accounts (NIPA), 44 National Institutes of Health, 354 National Institute of Justice (NIJ), 238, 244, 252-253 National Research Council, 236, 341, 365; Panel on Research on Deterrent and Incapacitative Effects, 102-103 National Science Foundation (NSF), 17-18, 39, 237, 242, 257; Law and Social Sciences Program, 253-254 "National Tax Test," 205 Negligence, 20 Netherlands, the, 54 Networks, social, 165-171 Newhouse, Joe, 341 New Mexico, 345 New York State Ice v. Liebmann, 348 Nicaragua, 360

Nixon, Richard M., 341 Noncompliers, prevalence of, 48—63 Nonfiling, 2, 19, 25, 89; age and, 134; education and, 118; prevalence of, 46, 50—51; in tax gap measures, 25, 43, 45, 47, 48 North Carolina, 136 Observability. See Visibility of income Occupational groups, 164-165, 167 Office of Management and Budget (OMB), 188 Ohio, 101 On American Taxation (Burke), 339 One-sided error component, 308-309, 310, 335 Ordinary least squares (OLS) estimation, 311-312, 313, 317 Oregon, 100-101, 128 Overreporting, 3, 25, 89, 95; capital gains, 26, 52, 60, 63, 70 n. 9, 107-108, 115; complexity and, 23, 59-63; defined, 21; prevalence of, 51-54; in research, 232; in statistical analysis, 319; tax practitioners and, 193 Panel studies, 49, 222, 226-227; of moral commitment, 122; statistical analysis of, 327-334 Panel on Taxpayer Compliance Research, ix, 18-20 Partnership income, 52, 108 Payer Master File, 29 Payne, John W., 153, 384, 390 Payroll taxes, 18 Peat Marwick Foundation, 252 Penalties: civil, 30, 31, 103-104, 227, 228; criminal, 31-32, 103-104, 111—112, 128; and deterrence, 86, 93—96, 106, 167; and moral commitment, 122; perceptions of, 130-133, 138, 212; probability of, 6-7, 27, 72, 81, 90, 110-112, 253; research and, 77-80, 83-84, 92-112, 152, 342-343; tax practitioner, 35, 176; TRA86 and, 16—17; in utility theory, 82, 83-84, 87, 88. See also Enforcement, tax law Pennsylvania, 164, 372; state tax, 175, 217 Pension income, 60 Perng, S.S., 241-242, 342, 349

Index Personal exemption, 183 Personal property taxes, 18 Philadelphia lawyers, 143 Poirier model, 326 Police Foundation, 358, 359 Political contributions, 59, 61 Poor, 147, 169, 184 Pornography laws, 370 Porter, Joan, 354 Preparers, tax. See Tax practitioners Presence effect. See Ripple effect President's Science Advisory Committee, 34i Prevalence of noncompliers, 48-63 Principal-agent model, 85 Privacy, taxpayer, 54, 171, 208-209, 222-223, 234-236, 238-239, 257-258, 343, 349-354 Privacy Protection Study Commission, 225-226 Private foundations, 250-252 Proactive contacts, 198, 202 Project 778 (IRS), 12, 13, 140 n. 13, 208, 210-211, 212; updating, 227-229, 257 Proprietor's income, 52, 53, 59, 60, 107-108 Psychologists, 77, 89, 124 Psychology, fiscal, 123; social, 149, 153-155 Public Health Services Act, 352 Punishment. See Penalties Quasi-experiments, 356, 367-370 Racc, 81, 133, 136-137 Random-audit assumption, 103-104 Randomized experiments, xi, 209, 211, 229-234, 243; defined, 340; feasibility of, 346-360; previous research, 341-346 Randomized response technique, 224, 225-226, 245 n. 7, 334-335, 343-344, 350-351 Reactive contacts, 198, 202-204 Reactivity, 362-364 Real estate taxes, 61 Record keeping requirements, 22-23, 26, 37, 87, 14« Regional variations in compliance, 163, 187 Regression-discontinuity experiments, 368-369 Regulatory laws, 147-148

399

Rehabilitation, 93-94 Remittance gap, 19, 41, 42, 43, 47 Rent income, 26,52,53, 108 Republicans, 126 Research: on compliance costs, 118; deterrence, 78—79; on income, 116-117; I R S and, 37-39, 67-69, 209, 210-213; on I R S and tax reform, suggested, 186-190; methodology, 220, 334-335; microeconomic theories, 80—90; on moral commitment, 118-133; on penalties, 92-112; politics of, 356-358; potential sponsors of, 250—256; quasi-experimental, 367-370; regression-discontinuity, 368-369; on social sanctions, 112-113; subjects, 213, 225-226, 231-234, 234-236, 241-245, 352-354; taxpayer, 19-20, 77-80, 181-186, 197-206, 248-250; on tax practitioners, suggested, 190-197, 252; on tax rates, 114-116; time-series, 369-370 Residential energy credit, 52, 61 Restaurant industry, 29. See also Tip income Retesting bias, 327-328, 333 Return-based compliance measure, 79, 232 Revenue agents, 30 Richardson, Aguayo v., 243, 348 Richardson, California Welfare Rights Organization v., 348 Riecken, Henry W., 341 Ripple effect (presence effect), 42, 103, 208, 212, 227-228; of audits, 6, 72, 103 Risk aversion, 83-84 Rivlin, Alice, 341 Rolph, John, 392 Rossi, Peter H., 341 Royalty income, 26, 52, 53, 108 Sales taxes, 18 Salk vaccine, 355 Sample selection bias, 317-318, 321 Sample selection model, 320-322, 323, 325 Sanctions. See Informal social sanctions; Penalties Sandmo, Agnar, 80-82, 83 Schedule C income, 30, 52; overreporting, 63; underreporting, 60, 107-108, 165, 218

4oo

Index

Schedule D. See Capital gains income Schedule E income, 26,52,59, 63, 165 Schedule F. See Farm income Schmidt, Peter, ix, 307 Schultz, Theodore, 360 Schwartz, Richard D., 384, 390 Schwartz-Orleans experiment, 78-80, 120, 124, 154, 155, 229-230, 242, 342, 367, 372 Scotchmer, Suzanne, 89, 90 Seat-belt use, 166 Self-employment: consistency and, 159; noncomplier prevalence in, 52, 54; in research, 367; and tax practitioner use, 174; underreporting in, 165 Self-reporting: biases in, 121-123, 136; commitment in, 125; and compliance measures, 222; distortion of, 211—212; equity and, 127-128; improving, 208; income and, 137; perceptions of evasion in, 120; social networks and, 112—113, 167. See also Surveys, taxpayer Sex, 81; correlation with compliance, 133, 135-136 Shoplifters, 160 Short form. See Tax forms Slemrod, Joel, 89-90, 118 Small businesses, 34, 367, 371 Smoking, 166 Social commitment. See Moral commitment Social Experimentation (Rjecken et al.), 341 Social networks and social fields, 9, 165-171 Social psychology, 149, 153-155 Social sanctions. See Informal social sanctions Social Science Research Council, 236, 341, 360 Social Security Administration, 45, 46 Socioeconomic status, 81, 137-138 Sociological theories, 81, 9 1 - 9 2 Sociologists, 77, 124 Spain, 130 Special Enforcement Program, 31-32 State dependence, 332-333 State and local taxes: deductions, and compliance in reporting, 52, 53, 59—63; rates, 114-115, 115-116; in research, 19, 1 0 0 - 1 0 1 , 175, 217, 249-250 Static random-audit models, 81, 83-84 Steel production, 164

Stochastic frontier error, 323 Stochastic frontier models, 308-316, 327, 335 Strategic noncompliers, 132 Substitution effects, 196 Subtraction items, 20; overstated, 2, 45, 134; visibility and, 107, 109. See also Business expenses; Deductions; Exemptions Survey of Individual Returns Filed ( T C M P Phase III), 12, 65-69, 175, 256 Surveys, taxpayer, 12, 13, 19, 213; on audits, 97—98; costs of, 36; as data base, 210, 211, 212; detection in, 96-97; and moral commitment, 119—123; and noncomplier prevalence, 49, 54—58; on perceptions of equity, 127-128; on practitioner use, suggested, 184, 191-192; and social sanctions, 112-113; statistical analysis of, 334-335; suggested improvements in, 208, 220-227; on T R A 8 6 , suggested, 185-186; underreporting in, 25-26. See also Research Sweden, 101, 131

Tax administration, 227-228; innovations in, 86, 198, 208; practitioners and, 190; and taxpayers, 154, 181-182, 197—199; and tax reform, 186-188. See also Enforcement, tax law Tax advisors. See Tax practitioners Tax auditors, 30 Tax careers, 156-162; research on, 208, 220-222 Tax Equity and Fiscal Responsibility Act ( T E F R A ) , vii-viii, 29; and T C M P , 215 Tax forms, 134; alternative, 344-345, 372; complexity of, 32-33, 118, 161, 197; confidentiality of, 234—236 Tax gap, vii, 1, 16, 2 4 - 2 5 , f i g . 43, 63; estimated components of, 19; income noncompliance, 41, 4 6 - 4 7 ; individual income reporting, 4 1 - 4 2 , 47; net, 47—48; unreported personal income, 41, 42, 4 3 - 4 6 Tax law. See Internal Revenue Code Taxpayer Compliance Measurement Program ( T C M P ) , 23-24; auditors, 215, 219; audits, 2, 49, 58, 66-68, 308,

Index 309, 316, 317; cost of, 35-36, 69-70 II. 6; data, 107, fig. 108, 210, 211-213, 309-316, 326-327, 327-329, 333, 335-336; demographics in, 133-134; discrepancy findings, 25, 26; effectiveness of, 139 n. 6; and enforcement security, 239-240; measure of unreported income, 45-46, 121; occupational groups in, 164; preaudit survey, 213; rcaudits, 95-96; research uses of, 67-69, 194-195; suggested modifications, 12-13, 208, 214-219; Survey of Individual Returns Filed (Phase III), 12, 65-69, 175, 256; tax practitioners and, 172,fip. 173, 214, 216; tax rates in, 115; TRA86 and, 183, 215, 217; V R P in, 65 Taxpayers, 21; audits and, 23-24, 71-72, 219, 327-328; chronic evasion, 159-160; decision making, 151—155; demographic characteristics, 133-138; in economic research, 85; environments of, 145-146; errors, 87, 88, 89-90; IRS contacts with, 28-29, 31, 197-206; and moral commitment, 73, 119-123, 125-126; occupations of, 214-215, 217-219; perceptions of tax system, 97-102, 104, 127, 128-133, 142-144, 153-154, 221; privacy of, 54, 171, 208-209, 222-223, 234-236, 238-239, 257-258, 343, 349-354; social influences on, 73, 162—171; and tax policy changes, 10, 181-190; tax practitioners and, 171-178, 190-193; and tax schemas, 149-151. See also Compliance costs; Self-reporting; Survevs, taxpayer Tax practitioners, 4, 10-11, 148, 345, 367; complexity and, 143-144, 174; ethics of, 193-194, 195-197; influences 011 taxpayers, 171-178, 191-193; IRS and, 27, 33, 34-35, 175-176, 203; regulation of, 27, 34-35; research on, suggested, 36, 87, 88, 190-197, 252; TCMP and, 172, 173, 214, 216; TRA86 and, 183-184, 192 Tax protesters, 31, 32, 130 Tax rates: bracket differential, 84, 139 n. 3; and compliance, 72, 114-116; in research, viii, 83-84, 85, 117; TRA86 and, 7, 116, 139 n. 3, 181, 184

401

Tax Reform Act of 1986 (TRA86), 7, 16, 112; and compliance costs, 118; and practitioner use, 183—184; 192; research potential of, 10, 181-190, 252; and tax gap, 47; and taxpayer confusion, 33; tax rates and, 7, 116, 139 n. 3, 181, 184; TCMP and, 183, 215, 217; and visibility, 30; defined, 149-151; development of, 155-171, 181-182; tax practitioners and, 171-178; TRA86 and, 184-185 Tax shelters, viii, 16-17; abusive, 31, 32, 187 Television: experiments, 359; IRS use of, 205 Time-invariant regressors, 331 Time-series experiments, 369-370 Tip income, 31, 46; noncomplier prevalence in, 52; reporting, 26, 41, 59, 60, 75, 92, 99, 108, 168; T E F R A and, 29 Tobin, James, 318 Tobit model, 318-320, 322 Total positive income (TPI), 66 Truax v. Corrigan, 348 Truckers, 167 Trust income, 26, 108 Tyler, Tom R., 129

Uncertainty, 82, no, 152, 234, 253; and economic research, 89-90 Underground economy, vii, 16, 24, 38, 83, 157, 167, 170; estimated size of, 41, 42; and unreported income, 43-46 Unemployment compensation, 60, 117 United Kingdom, 126, 130 U.S. Congress, 16, 188; Joint Committee on Taxation, 254-255, 371 U.S. Department of Health, Education, and Welfare, 353 U.S. Department of Health and Human Services, 241-242, 244, 347, 352-353 U.S. Department of Housing and Urban Development, 347 U.S. Department of Justice, 32, 252-253 U.S. Department of Labor, 214-215, 218 U.S. Department of the Treasury, 16; Circular 230, 35, 193-194; Office of Tax Analysis (OTA), 254-255 Unreported taxable personal income, 41, fijj. 42, 43-46 Utility, 80-82, 83, 89, 139 n. i, 175

402

Index

Vertical equity, 127 Vietnam War, 134 Visibility of income, 5, 26, 71, 81, 88, 217; computer records and, 228; and detection probability, 106—no; in deterrence research, 104; encourages compliance, 53; increasing, 27-30; research and, 218—219 Voluntary compliance level (VCL), 31, 58, 70 n. 12, 95; analysis of, 102—104; audit class and, 140 n. 8; defined, 68; income level and, 117; in Project 778, 227; race and, 136—137; socioeconomic status and, 137; by tax practitioner type, 173, 177; visibility and, 107 Voluntary income tax assistance (VITA), 33, 169 Voluntary reporting percentage (VRP), 26-27, 70 n. 13; measuring, 58-6$

Voting behavior, 126, 158-159 Wage income: noncompliance amounts, 107—108; noncomplier prevalence in, 52, 53; V R P , 60, 64, 65 Waitresses, 168 Wall Street Journal, The, 76 Watergate, 134 Watts, Harold, 341 W-4 form, 31, 189 Wilson, Joseph F., 341 Withholding, 71, 87, 88, 134; employers and, 43; income and, 137; interest, 37; taxpayers and, 151-152; and visibility, 26, 107; and V R P , 59; wage, 28. See also Information returns Women, 135-136; wages of, 320-321 Work/Welfare Demonstration, 357 W-2 form, 75